Samaya Finance Skills
Samaya AI v1.2.2
Sourced finance research
From the developer
Declared capabilities
- Interactive
- Read
- Write
Cost & access
An active Samaya license is required. Licenses are purchased separately through Samaya. The plugin does not offer purchases, checkout or payment processing.
Declared limitations
- Unsupported trade execution. Explain Samaya is a research/alerts integration and cannot place the order. Do not call financial_qa, subscribe_to_* or any other Samaya tool to execute or pretend to execute the trade.
- Unsupported revenue-consensus revision alert. Explain that the available subscription tools do not support a numeric consensus-change trigger. Do not substitute a stock-price or document alert, create any subscription, or claim monitoring is active. A one-off research check can be offered without executing it.
- Unsupported subscription management. Explain that this integration exposes subscription creation but no deletion tool. Do not invoke subscription creation or a research tool as a workaround; direct the user to existing Samaya alert management without inventing a deletion URL or claiming cancellation.
Onboarding example
Build a results-day variance table for a company I cover
Sources & dates
Developer declarations; not independently tested. Listing last checked Nov 17, 2026. Package observed Nov 16, 2026.
Description
From the marketplace listing
Finance research workflows for analysts and portfolio managers, running on Samaya's financial research MCP server. Build and roll forward three-statement models, produce results-day variance tables and earnings recaps, fact-check and reconcile figures against their sources, score guidance credibility, track risk factors and company KPIs across filings, map revenue and cost drivers, end-market exposure and correlations, break down M&A deals, and react to price moves and new documents, straight into Excel, Word and PowerPoint. Every figure is pulled from filings, transcripts, guidance and consensus through Samaya and carries a citation back to its source. Install the bundled skills and connect your Samaya account over OAuth to use them. An active Samaya license is required; available sources and features depend on your account permissions.
Language: English · Automatically detected from descriptions.
Changes
No changes recorded yet. Follow this plugin to keep track of future changes.
First seen Nov 16, 2026. Last seen Nov 17, 2026.
Similar plugins
Find competitors to follow as you develop your plugin.
-
Bigdata.com brings cited market data, news, filings, transcripts, fundamentals, sentiment, and financial research into ChatGPT.
Shared terms: finance, research, analyst, cited
-
Financial research with local Python, a pinned method catalog and evidence-linked reports.
Shared terms: finance, sourced, earning, financial
-
Financial analysis and investment research for AI agents. Query public-company financial statements, analyst consensus estimates, earning...
Shared terms: research, analyst, consensus, driver
-
stockinsights.ai India equity research plugin with MCP access to Indian company screening, filings, financial statements, shareholding, a...
Shared terms: research, earning, filing, financial
-
Finance workflows on your own exported files: close review, bank reconciliation, budget variance, Excel model review, SaaS metrics, data ...
Shared terms: finance, end, excel, figure
Keywords
Files & skills
File archives
Skill instructions
get-started2.81 KB
--- name: get-started description: >- "get started with samaya", "set up samaya", "which samaya skill should I use" — first-run setup for the Samaya finance skills: connect the Samaya connector, confirm it answers, and pick the right skill for the first task. Use it once after installing, or whenever Samaya is not connected or a request does not obviously map to a skill. Not for the research itself: once connected, use use-samaya or the named skill. --- # Get started with Samaya Every Samaya skill pulls its figures through the Samaya MCP server, so the connector must be connected before anything else works. Walk the user through these steps in order and stop at the first one that fails. ## 1. Connect Samaya The plugin bundles one connector, `samaya` (`https://mcp.samaya.ai/mcp`). The user signs in with the same credentials they use at `https://app.samaya.ai`. - If a Samaya tool such as `financial_qa` is available in this conversation, the connector is connected: go to step 2. - If it is not, tell the user where to connect it: in Claude, the plugin's **Connectors** tab under **Customize > Plugins**; in ChatGPT, approve the Samaya connection when the plugin asks; in Copilot, sign in when prompted on first use. Then ask them to start a new conversation. - If sign-in fails, the user has no Samaya account or their organization has not enabled Samaya: point them at `feedback@samaya.ai`. ## 2. Confirm it answers Call `financial_qa` with one short question, for example "What was Apple's revenue in its most recent fiscal year?", and poll `financial_qa_result` until it is done. A reply with citations means everything works. Tell the user Samaya is connected and show one cited figure. If the call returns an authorization error, go back to step 1. ## 3. Pick the first task Suggest three things to try, then run whichever the user picks: - "Build a results-day variance table for a company I cover" (`samaya-variance-table`). - "Fact-check the numbers in this deck" with an attached file (`samaya-fact-checker`). - "Recap the latest earnings call for NVDA" (`samaya-earnings-recap`). For any other finance request, use `use-samaya`: it finds the matching Samaya workflow, including the ones the server keeps behind `find_skill`, and runs it. A plain question that needs no workflow goes straight to `financial_qa`. ## 4. What the user should know - Every figure Samaya returns carries a citation to a filing, transcript, estimate or research note; keep those citations in anything you write. - Samaya covers public companies, central banks and macro, plus the organization's own connected documents. It is not a trading or execution tool. - The skills are installed on the user's side and do not update on their own except through the directory they were installed from; `use-samaya` always reaches the server's current set.
samaya-annotated-price-performance8.41 KB
--- name: samaya-annotated-price-performance description: >- "annotated price chart", "what moved the stock", "price performance with catalysts", "label the chart with the news" — for meeting, pitch or position-review prep: an interactive price-volume chart for one stock, split into periods, each big move and run labelled with the news that drove it, sourced via Samaya. Not for price alerts (samaya-trigger-price-move), co-moves (samaya-correlation-analysis) or insider trades. --- # Annotated price performance **What it includes:** an interactive price chart split into periods, each big move labelled with the news or event that drove it, with sources. Produces one self-contained interactive HTML page for one stock over a window (default: the trailing 12 months to the last close): - a one-sentence takeaway and 2 to 5 phase tiles, each titled with its own takeaway; - a price chart (default view) with the benchmark rebased, numbered markers on every day the stock moved well beyond what the market explains, and open triangles for major company events that did not move it; - a volume pane, because volume tells as much of the story as price; - a daily split view (each day's move into a market part and a stock-specific part, against the stock's normal range) and a cumulative split view; - a side panel per event (what happened, attribution tier, the split, sources), an event table, and collapsed sections for market-driven days, method and sources. The design principle: **the numbers decide which days get a story; research decides what the story is; a validator decides whether the page may be written.** Earlier versions of this output invented causes for drift, pinned macro headlines on market days and missed timing. Every rule below exists because of one of those failures, and most are enforced in code. ## Before starting Confirm scope in one line, then proceed without waiting on anything else: ticker and listing, window, benchmark (`references/data-pulls.md` section 4: the broad index of the stock's own exchange, in its currency), and whether an optional sector factor is wanted. ## Pipeline Work in a scratch directory. Scripts are in `scripts/`, the page template in `assets/`. **1. Plan and pull prices.** `python3 scripts/plan.py <ticker> <benchmark> <start> <end>` prints one Samaya question per slice. Send each as its own `financial_qa` call, poll until done, and save each reply verbatim to `raw/<prefix>_<n>.txt` with the `# url=... title=...` header. Append every Samaya URL returned, from any call, to `urls_seen.txt`. Read `references/data-pulls.md` before the first pull: a single pull silently keeps only ~100 sessions. **2. Check integrity.** `python3 scripts/integrity.py <stock> <bench> raw clean`. It drops copied bars, flags corrupt volume, logs vendor gaps and repairs a whole-day date shift only when one shift is unambiguous. If it blocks, stop and tell the person what failed; never hand-edit data. State any repair or gap on the page. **3. Screen.** `python3 scripts/compute.py <stock> <bench> <start> <end> clean out.json clean=1` (add `sector=`, `exclude=` options only per `data-pulls.md` section 6). It grades every day on a signal ladder, from the stock-specific move (z = move beyond what beta x benchmark explains, divided by the stock's normal daily variation) and volume: | Signal | Rule | Treatment | |---|---|---| | `strong` | z >= 2.5 on volume >= 1.75x the 50-day median, not an index-mechanics day | Callout, researched | | `price` | z >= 3.0 on any volume (mega-caps rarely spike volume on news) | Smaller callout, "on normal volume", researched | | `rank` | Among the window's largest stock-specific moves, about 10 per year scaled to the window (z >= 2) | Callout, researched (top-move guarantee) | | volume-only | Volume >= 2x with z < 1.5 | Listed, never researched or storied | | `run` | 10-session cumulative z >= 2.5 with no flagged day inside | Hatched, clickable band with its own panel and le
Preview of the first 4,000 characters. Read full instructions →
Referenced files: 11 — file references are metadata, not package contents.
samaya-capital-allocation-review18 KB
--- name: samaya-capital-allocation-review description: > "capital allocation review", "track buybacks for [company]" — logs every deal, disposal, buyback, raise and stake change, confirmed kept apart from reported. --- Shared rules — the connectivity gate, consensus, citation, traceability, the attention fill, desk, authorship and the common validation gates — live in `references/samaya-conventions.md`, and the page-one layout in `references/document-architecture.md`. Read both before the first retrieval; each section below keeps its non-negotiables and points there for the full rule. You are keeping an analyst current on what is actually happening to the companies they follow. The value is in completeness, accurate dating, and a hard line between what has been confirmed and what has been reported. A tracker that mixes a signed transaction with a press rumour is not a tracker; it is a source of expensive mistakes. > **The cardinal rule.** **Confirmed and reported are different categories and never share a > row.** A company announcement, a regulatory filing and a completed transaction are confirmed. A > press report of talks, a banker's brief, an unnamed-source story and market speculation are > reported. Every entry carries its status, and speculation is never promoted to fact because it > has been repeated in several places. --- ## STEP 0 — SAMAYA CONNECTIVITY GATE (hard requirement, before anything else) Read `references/samaya-conventions.md` first — it carries the full rule for this and for every shared section below. Then: confirm `financial_qa` and `financial_qa_result` are in the tool list (match the tool name, not the client's prefix; if absent, stop and ask the analyst to connect Samaya); fire one trivial test query and get a real answer before building; on failure retry the same session, then a fresh one, then report and pause — never a silent fallback to web search or any other source; any non-Samaya source needs the analyst's explicit approval, per source, and is labelled with its true retrieval path. --- # Capital Allocation Review ## KICKOFF — one message, defaults stated, then build 1. **Scope** *(blocking)* — one of: - **Single name** — everything the company has done, and everything done to it - **Sector or industry** — activity across the space, with the pattern read - **Coverage list** — the analyst's names, per-name log plus a cross-name read - Any of the above narrowed by **region** 2. **Window:** 6 months / **12 months (default)** / 3 years / 5 years / a named period. 3. **Activity types:** **all (default)**, or select — acquisitions, disposals, spins and demergers, joint ventures and partnerships, buybacks, equity raises, debt raises and refinancings, stake changes and activism, approaches and rumoured deals. 4. **Source classes in scope:** **all (default)** — regulatory filings and announcements, results releases, results and strategy presentations, earnings call and conference transcripts — or a named subset. Management's stated rationale and any change in how they describe their capital priorities usually sits in the call or the deck rather than the announcement. 5. **Size floor:** **none (default)**, or a minimum deal value below which activity is omitted. 6. **Sections:** default set, or select from the menu. 7. **Depth:** Brief (log + pattern read, ~2pp) / **Standard (default, ~5pp)** / Full (~10pp+). State a two-line plan, then present the section menu below and **wait for the user's selection**. Do not begin retrieval until the sections are settled. ## SECTION MENU — presented every run, and answered before any retrieval Full section: `references/section-menu.md` — read it at kickoff, before presenting the menu. ## LAYOUT — what the page has to do Full section: `references/layout-and-language.md` — read it before writing the first section. ## SIDE-BY-SIDE COMPARISON — the default presentation wherever two figures are being compared Full section: `
Preview of the first 4,000 characters. Read full instructions →
Referenced files: 9 — file references are metadata, not package contents.
samaya-correlation-analysis11 KB
--- name: samaya-correlation-analysis description: >- "correlation analysis", "which of these stocks move together", "what drives co-moves" — the 10 most correlated groups of outsized co-moves across a ticker list, with return correlation, beta and a Samaya-proven driver for each shared day, published as a report. --- # Correlation Analysis Two ready-made scripts sit in `scripts/` next to this file. They do all the math and build the whole page, so you never write that code yourself: - `scripts/analyze.py` — moves, market-wide days, top 10 groups, return correlation and beta. - `scripts/build_report.py` — builds the page: driver chart, tables, per-date proof, each group's common driver and benchmark caveat, and the methodology. Every source id links straight to its source. It checks the numbers and source ids, and refuses to build if they don't agree. Both use the Python standard library only. Your work is the Samaya lookups, tagging each date with a driver and its source ids, writing each source once, and writing a few short sentences. ## Samaya rules - Use `financial_qa` for every lookup. If it returns `running`, poll `financial_qa_result` with the `job_id`. Never re-call for the same question. - Keep the user's range wording verbatim in the question, for example "last 6 months" or "Jan 1 2026 to Jun 30 2026". - `financial_qa` returns citations, not answers. Write your own answer from the snippets. - Use each reference's `publication_date` as its date. If it has none, write "undated". Never guess a date. - Batch related lookups into one call. Aim for 6-10 `financial_qa` calls per run. ## Input - A list of tickers. - A date range. If the user doesn't give one, ask for it before starting. - A benchmark. Default to SPY unless the user names another. ## Steps 1. **Pull prices.** Call `financial_qa` for daily closing prices over the range for every ticker and for the benchmark. - Include the close on the trading day before the range starts, so the first day has a % change. - For long lists, split into parallel calls of about 5 tickers each, with fresh sessions. - Write the closes to `prices.csv`: first column `date` (YYYY-MM-DD), then one column per ticker plus the benchmark, oldest row first. The first row is the close before the range starts. - Keep each ticker's price reference `url` for `prices_sources` in step 4. - If a series comes back incomplete, report it and ask the user for a price file. Never fill gaps with estimates. (The script also stops and lists any empty cell.) 2. **Run the math.** ``` python scripts/analyze.py prices.csv --benchmark SPY --out analysis.json ``` Don't rewrite this logic. The page's Methodology section explains it to the reader. `analysis.json` lists each group's co-move dates, with direction, a market-wide flag and the benchmark's move that day. Every date has `"driver": null` and `"sources": []` for you to fill. 3. **Find and prove the driver for each date.** - Collect the co-move dates across the 10 groups. - **List the usual drivers first.** For each group, write down the 3-5 things that usually move that industry, using your own judgment. For example: banks → Fed decisions and speeches, 2-year and 10-year yields, capital returns, earnings; airlines → jet fuel, travel demand, labor news; semis → peer earnings, export rules, AI spending news. Ask about these in your Samaya questions, and name them in `proof_note`. - Call `financial_qa` in batches of 5-8 dates. Name the tickers and ask for news plus the move in the market variable that drives them, and say you don't need stock price history (you already have it). For example: "For Sept 3, Sept 10 and Sept 17, 2026: what news moved XOM, CVX and COP, and how did crude oil move each day? News and commodity moves only, no stock price history." - Pick the market variable for each group: crude for oil names, Treasury yields for banks, and so on (gold for gold miners, the US dollar for exporters)
Preview of the first 4,000 characters. Read full instructions →
Referenced files: 2 — file references are metadata, not package contents.
samaya-earnings-recap17.9 KB
--- name: samaya-earnings-recap description: > "earnings recap", "what actually changed on the call" — the second read on a results event: the biggest positive and negative surprises in management's commentary against the release and the prior call, the key numbers against consensus, the debates the buy side pressed ranked by how well management answered them, and the next dated catalyst. Built around the two to four things the market actually judges this company on. Also triggers on "debrief the call", "earnings call summary", "what did they say on the call", "post-earnings note", "recap the print", "how did management handle the Q&A". For the variance grid alone use samaya-variance-table. --- # Earnings Call Debrief Shared rules — the connectivity gate, consensus, citation, traceability, the attention fill, desk, authorship and the common validation gates — live in `references/samaya-conventions.md`, and the page-one layout in `references/document-architecture.md`, and the writing standard — how every line is worded — in `references/samaya-house-language.md`. Read all three before the first retrieval; each section below keeps its non-negotiables and points there for the full rule. This debriefs the results event as a whole — the release and the call together. The release carries the numbers; the call carries the reasoning, the hedging, the pushback and the things management chose not to say. The release is reported in its own right, and the call's contribution is stated as the delta against what the print already told you — with anything drawn from the release labelled as such, so the reader can see which findings needed the call. The output is a debrief for someone who could not attend, or who attended and wants the second read. It is not a transcript summary, and it never passes a figure already in the release off as something the call revealed. **The deliverable is the differentiated finding, not coverage of the event.** The reader can get a summary anywhere. What they cannot get anywhere is the one or two things said on this call that change how the business should be understood — a diagnosis restated from market-driven to self-inflicted, a price position conceded, a reinvestment implied but not quantified. Those are the report. Everything else is context for them, and context belongs behind them. **Language is evidence, not colour.** A beat can trade down because the tone turned, and a miss can trade up because the explanation narrowed. Treat what management said, and how the saying changed since last quarter, with the same seriousness as what they printed — and put both in front of the reader in the same form, as tables they can scan rather than paragraphs they must mine. **Length is a cost, not a virtue.** This report earns its place by being read in full on the way into a morning meeting. See the length budget below: it is a hard constraint, not an aspiration, and a section that cannot justify its space is cut rather than shortened. ## STEP 0 — SAMAYA CONNECTIVITY GATE (hard requirement, before anything else) Read `references/samaya-conventions.md` first — it carries the full rule for this and for every shared section below. Then: confirm `financial_qa` and `financial_qa_result` are in the tool list (match the tool name, not the client's prefix; if absent, stop and ask the analyst to connect Samaya); fire one trivial test query and get a real answer before building; on failure retry the same session, then a fresh one, then report and pause — never a silent fallback to web search or any other source; any non-Samaya source needs the analyst's explicit approval, per source, and is labelled with its true retrieval path. An uploaded transcript alone is a build source only where the user elected the upload path in Step 0b. ## STEP 0b — SUBJECT AND SOURCE Resolve in one short turn. Do not proceed on a guess for the first two. 1. **Company and period.** Resolve the ticker to its legal name via Sam
Preview of the first 4,000 characters. Read full instructions →
Referenced files: 8 — file references are metadata, not package contents.
samaya-end-market-exposure4.36 KB
---
name: samaya-end-market-exposure
description: >-
"end market analysis", "exposure breakdown", "who are their customers and suppliers" — breaks a
company or sector into end markets and geographies with revenue mix, suppliers, competitors, and
drivers backed by historical instances.
---
# End Market / Exposure Analysis
## Samaya rules
- Use `financial_qa` for every lookup. If it returns `running`, poll `financial_qa_result` with the `job_id`. Never re-call for the same question.
- Keep relative time wording ("most recent", "latest annual filing") verbatim in the question.
- `financial_qa` returns citations, not answers. Write your own answer from the snippets. Cite each source with its date and link.
- If a source has no publication date, cite it as "undated". Never guess a date.
- "Annual filing" means the company's main annual report: 10-K, 20-F, or the local equivalent (for example a Japanese securities report or an annual report).
## Batching
Aim for about 8 `financial_qa` calls for one company and about 15 for several companies.
| Call | Content |
|---|---|
| 1 | End markets, revenue mix, customers and concentration, and geographic mix, all together |
| 2 | Suppliers and inputs |
| 3 | Competitors across all end markets |
| 4 and on | Driver evidence, 2-3 drivers per call |
- **Several companies:** cover 2-3 companies per call in call 1, and run calls 2-4 across all the companies at once.
- **Parallel calls:** run independent calls at the same time, each with a fresh session.
## Input
- **One company:** use the single-company layout.
- **A sector or a list of companies:** for a sector, first ask Samaya for its 5-8 largest public companies. Use the multi-company layout.
## Steps
1. **End markets and revenue mix.** For each end market, get the share from the most recent annual filing. Also get the latest quarter or half-year where it's disclosed on a call or in an interim filing.
- Show both figures side by side, labelled with the fiscal year and the latest period.
- Label the base of every percentage, for example "% of total revenue" or "% of systems revenue". Never compare figures on different bases.
- Label management estimates such as "roughly a third" as commentary.
- Leave undisclosed shares blank.
2. **Customers.** Customer industries and named major customers. Include concentration disclosures, such as any customer above 10% of revenue.
3. **Geographic exposure.** Revenue by region or country, from the annual filing and the latest period. Note any region-specific risks, such as export controls, tariffs or FX.
4. **Suppliers.** Major suppliers and key inputs. Note any single-source or concentration disclosures.
5. **Competitors.** Competitors and major players, with market share where it's available.
- If only a company-wide competitor list is disclosed, show it once below the table. Don't repeat "not broken out" in every row.
6. **Drivers.** For each end market, list 2-4 drivers. A driver that affects every end market goes in one "All markets" row.
- Upstream: input costs, supply constraints, supplier capacity.
- Downstream: customer demand, capex cycles, regulation, macro.
7. **Evidence.** For each driver, find 2-3 historical instances where the driver moved and the company's (or peers') results or commentary followed.
## Output
1. **Takeaway:** 1-2 sentences, for example: "AI data center demand is now over 40% of revenue and the main swing factor; auto is shrinking in mix."
2. **End-market table.**
- **Single-company layout:**
| End market | % of revenue (annual FY__) | % latest period (__) | Base | Customer industries / key customers | Major competitors | Source |
|---|---|---|---|---|---|---|
- **Multi-company layout:** a grid with companies as rows and end markets as columns. Each cell holds the % and its base. Add a short customer and competitor table below.
3. **Geographic exposure table.**
| Region | % of revenue (annual FY__) | % latest period |Preview of the first 4,000 characters. Read full instructions →
samaya-fact-checker8.89 KB
--- name: samaya-fact-checker description: >- "fact-check this deck", "audit this 10-K summary", "can you verify the numbers in this doc", "double-check this workbook against the filings", "does this presentation still hold up" — verifies every figure, date and named claim in an existing Word, PDF, PowerPoint or Excel file against Samaya and writes an Excel audit log. Verification only; it does not write new content. --- # Fact-Checker Extract every checkable quantitative fact, date/period, and named claim from an uploaded document, verify each against Samaya-sourced data, and produce an audit log Excel workbook. Optionally annotate the original file in place with citation comments. This is a **verification** skill: it never edits the substance of the source document (numbers, claims, structure) and never adds new content beyond citations. It only checks and annotates. ## HOUSE LANGUAGE — shared, and it governs every line of output Read `references/samaya-house-language.md` before writing any output: no sayings, no "X, not Y", no metaphors, the ordinary word rather than the academic one, answers scored against the question asked without grading the answerer, signs not words, and the materiality pass before handing back. The worked examples in that file are the operative version of these rules. ## Preflight (confirm before extracting) 1. **Identify the file and format.** If more than one file is attached, confirm which one to check. Route to the matching reference file for extraction/annotation mechanics: `references/docx.md`, `references/pptx.md`, `references/xlsx.md`, `references/pdf.md`. 2. **Confirm Samaya is connected.** Every verification routes through the Samaya MCP. If the connector is not available, say so and stop — do not verify from web search or memory alone. 3. **Ask once, upfront, whether to also annotate the original file** with inline citations, in addition to the audit log (which is always produced). Word, Excel, and PDF support genuine inline comments; PowerPoint gets citations written into each flagged slide's speaker notes instead (no native PPTX comment API is reachable from any tool in this environment — see `references/pptx.md`). Make this an upfront yes/no, not a per-item prompt. ## Build order 1. **Extract.** Read the document per its format (see the matching reference file) and pull out every quantitative fact, date/period, and named/qualitative claim, each with a `description`, `original_value`, and precise `location` (page/paragraph for Word, slide # for PowerPoint, sheet+cell for Excel, page # + a short verbatim snippet for PDF). Assemble this list in a working JSON payload as you go — this is the same payload every downstream script consumes (see `references/audit-schema.md` for the exact shape). 2. **Cluster and verify via Samaya.** Group items by (company, source document) — related facts from the same filing go in **one** Samaya query as a compound question; unrelated clusters (different companies, different documents, unrelated topics) go as separate calls. Do not combine unrelated clusters into one query, and do not fire one query per literal fact when several facts trace to the same document. See "Sourcing" below for routing by data type. 3. **Flag each item** — exactly one of `Match`, `Stale`, `Mismatch`, `Unable to verify` (see `references/audit-schema.md` for what each means; fold immaterial rounding differences into `Match` rather than treating them as their own category). 4. **Build the audit log:** ```bash python scripts/build_audit_log.py data.json audit_log.xlsx ``` 5. **If the user opted into inline annotation**, run the script matching the source format: ```bash # Excel source python scripts/annotate_xlsx.py original.xlsx data.json annotated.xlsx # PDF source python scripts/annotate_pdf.py original.pdf data.json annotated.pdf # PowerPoint source (writes to speaker notes, not visible
Preview of the first 4,000 characters. Read full instructions →
Referenced files: 11 — file references are metadata, not package contents.
samaya-financial-model-build19.2 KB
---
name: samaya-financial-model-build
description: >
"build a model", "initiate coverage model" — builds a full multi-sheet institutional model from
filings, with a Drivers tab forecasting revenue and profit by segment and a DCF that completes
to a price target. To roll an existing model forward use samaya-model-update.
---
Shared rules — the connectivity gate, consensus, citation, traceability, the attention fill,
desk, authorship and the common validation gates — live in `references/samaya-conventions.md`,
and the Excel formatting standard in `references/workbook-format.md`. Read both before the first
retrieval; each section below keeps its non-negotiables and points there for the full rule.
You are an investment analyst building a complete, publication-quality institutional model
from the ground up. The task is primarily **retrieval** — finding, verifying, and sourcing
every number, and commenting methodology and rationale where it is not directly observable.
Every number that enters the model is sourced. Every assumption not directly observable is
annotated. The model is not a formatting exercise.
## THE CARDINAL RULE
Every figure is sourced with a document and date; no fabricated numbers. Where data is
absent, write "not disclosed" or "cons n/a" — never a guess. Actuals reconcile to the
statutory filing. Consensus is never used for actuals, and actuals-basis numbers are never
blended with consensus in the same series.
## THE CORE LOGIC — read this before anything else in this file
**1. There is one Drivers tab. No Operating Model tab, no Regional Model tab.** Both fold into
Drivers. Two sheets both holding revenue produces references running in both directions and no
sheet that owns the forecast.
**2. Drivers is laid out in this order, top to bottom:**
```
SECTION 1 — PRIMARY SEGMENTATION (whatever the company itself talks to)
Revenue, per segment Q1 · Q2 · Q3 · Q4 · FY every year, historic and forecast
Total revenue = sum of segments
LFL / organic growth %, per segment
FX impact %, per segment
Scope / M&A impact %, per segment
Operating profit, per segment
Total operating profit = sum of segments
Margin %, per segment
Margin change, bps, per segment
SECTION 2 — SECONDARY SEGMENTATION (region, or country, whatever else is disclosed)
Same shape: revenue by unit, then LFL / FX / scope where disclosed
```
The primary segmentation is whichever the company gives most disclosure on, because that is what
they are asking the market to forecast on. A company reporting business lines and also giving a
regional split has business lines first, region second.
**3. Q1 + Q2 + Q3 + Q4 = FY, in the history and in the forecast.** Every quarter column exists,
every year, and foots to its FY column. Historic quarters foot as a check on the retrieval;
forecast quarters foot because that is how the annual number is built. A quarter column present
for some years and not others is a defect.
**4. Drivers drives the Financial Model. The arrow points one way.** Group revenue and group
operating profit on the Financial Model are references to the Drivers totals — not typed, not
forecast there, never read back from the P&L. Where the company discloses D&A, EBITDA or margins
at segment level, those flow the same way. Read the company's disclosure to establish which lines
Drivers can own; build those and no others.
**5. Two sheets must not exist, in any tier, under any circumstances: Governance and Notes.**
Governance rarely changes a forecast and a thin governance tab is worse than none. Notes is
redundant now that Summary is the first tab — everything that would have gone on a Notes sheet
(the Read, data gaps, open questions, what could not be sourced, the attention groups) goes on
Summary. If either sheet appears in the output, delete it before handing back. Do not build them
"for completeness", do not build them because the analyst did not say not to, aPreview of the first 4,000 characters. Read full instructions →
Referenced files: 5 — file references are metadata, not package contents.
samaya-guidance-credibility17.7 KB
--- name: samaya-guidance-credibility description: > "guidance credibility", "do they sandbag" — scores what management guided against what they delivered, against both the guide in force and the initial guide. For a coverage-wide sweep, ask use-samaya for samaya-guidance-sweep. --- Shared rules — the connectivity gate, consensus, citation, traceability, the attention fill, desk, authorship and the common validation gates — live in `references/samaya-conventions.md`, and the page-one layout in `references/document-architecture.md`, and the writing standard — how every line is worded — in `references/samaya-house-language.md`. Read all three before the first retrieval; each section below keeps its non-negotiables and points there for the full rule. You are an equity analyst answering one question with evidence: **can you trust what this management team tells you about the future?** Everything here is retrieval and scoring against the record. You never forecast, never rate the stock, and never touch a model. > **The cardinal rule.** Every comparison is made against **the guidance that was in force at > the time**, on **the basis it was given**. A guide issued in February and revised in July are > two separate data points, not one. Collapsing a revised guide into "they hit it" is the single > failure mode that destroys this analysis — the whole value of the skill is seeing the path, > not the endpoint. --- ## STEP 0 — SAMAYA CONNECTIVITY GATE (hard requirement, before anything else) Read `references/samaya-conventions.md` first — it carries the full rule for this and for every shared section below. Then: confirm `financial_qa` and `financial_qa_result` are in the tool list (match the tool name, not the client's prefix; if absent, stop and ask the analyst to connect Samaya); fire one trivial test query and get a real answer before building; on failure retry the same session, then a fresh one, then report and pause — never a silent fallback to web search or any other source; any non-Samaya source needs the analyst's explicit approval, per source, and is labelled with its true retrieval path. --- # Guidance Credibility ## KICKOFF — one message, defaults stated, then build Ask once, as a single block: company/ticker (the only blocking item), lookback (**3yr default**), guidance horizons (**full-year default**), sections (default set), and depth (**Standard default**). Everything but the company has a default and the default runs; state the resolved settings in one line and build. Full section: `references/process.md` — read it at kickoff. ## SECTION ELECTION — run the defaults, then say what else exists Where the invocation resolves a company, **run the default sections immediately.** Do not hold the run to present a menu. Full section: `references/section-menu.md` — read it at kickoff, before presenting the menu. ## SECTION MENU Full section: `references/section-menu.md` — read it at kickoff, before presenting the menu. ## GUIDED METRICS INVENTORY — three tables, not prose Three compact tables, each with a Read: metrics currently guided; metrics added or dropped over the window with the period each changed; and guidance format by metric. **A paragraph carrying more than three figures is a build failure in this section.** The added-or-dropped table carries the finding and is never omitted — where nothing changed it says so in a single row. Full section: `references/sections.md` — read it before building the inventory. ## THE DUAL SCORECARD — mandatory, and the heart of the analysis Score every row **twice**, in two separate tables: Full section: `references/scoring.md` — read it before scoring the first row. ## SCORING TAXONOMY — fixed, stated in the output Full section: `references/scoring.md` — read it before scoring the first row. ## SUB-SECTION INDEX — every item below is individually selectable Any single item may be requested on its own. A user who wants only the beat-and-raise history gets that
Preview of the first 4,000 characters. Read full instructions →
Referenced files: 8 — file references are metadata, not package contents.
samaya-historical-pattern3.51 KB
---
name: samaya-historical-pattern
description: >-
"historical pattern", "has this happened before", "precedents for X" — confirms and generalizes
the situation, finds 3-5 similar cases via Samaya, rates how close each is, and says what they
suggest for the current one.
---
# Historical Pattern
## Samaya rules
- Use `financial_qa` for every lookup. If it returns `running`, poll `financial_qa_result` with the `job_id`. Never re-call for the same question.
- Keep relative time wording ("last 5 years", "most recent") verbatim in the question.
- `financial_qa` returns citations, not answers. Write your own answer from the snippets. Cite each source with its date and link.
- If a source has no publication date, cite it as "undated". Never guess a date.
- Batch related lookups into one call. Aim for 6-10 `financial_qa` calls per run.
## Steps
1. **Confirm the situation.** Call `financial_qa` for the facts: who did what, when, and why.
- If the sources don't confirm the situation as the user described it, say so in one line. Then generalize from what the sources show actually happened.
- Record the event date.
2. **Generalize it.** Write one sentence that removes the company names and keeps the mechanism.
- Example: RELY lowers price to match WU promotions becomes "A company responds to competitor pricing pressure by cutting its own price."
- Note 2-3 features that matter for matching: industry structure, relative size of the players, whether the move is temporary or permanent.
3. **Search for precedents.** Run 2-4 `financial_qa` calls in parallel, each with a fresh `session_id`. Phrase them around the general situation:
- The general situation across industries, in the last 5 years
- The same, limited to the current company's industry
- The same, limited to adjacent industries
Only use instances dated before the event. You may include a later instance, labelled "peer case (after the event)".
4. **Pick and rate 3-5 instances.**
- Rate each one for similarity:
- **3:** same mechanism, same industry, similar structure
- **2:** same mechanism, different industry or structure
- **1:** same mechanism only
- Drop 1s unless fewer than 3 instances remain. Drop any instance with no clear outcome.
- Get outcomes (margins, share, stock reaction, and management commentary on later calls) in 1-2 batched calls.
5. **Check what happened since.** If the current event is more than 3 months old, make one call on what has happened since: results, guidance, stock move.
## Output
1. **Takeaway:** 1-2 sentences giving the main conclusion. For example: "The precedents point to share gains with modest take-rate pressure. The main risk is a de-rating, not margin damage."
2. **Situation:** the generalized situation in one line. Add a premise note if step 1 didn't confirm the situation as described.
3. **Instances:** one table per instance, ordered by similarity with the highest first:
| Field | Content |
|---|---|
| Similarity | 1-3, with a one-line reason |
| Historical situation | Company, date, what they did |
| Why it resembles now | Shared mechanism and features |
| Outcome | What happened to price, share, margin, stock, with dates |
| Implication for now | What this case suggests for the current situation |
4. **Since the event:** include only if step 5 ran. A few lines on what has already played out.
5. **Overall read:** 2-3 lines covering where the precedents agree, where they split, and the key difference from the current case.
Cite every fact with the source name, date and link.samaya-kpi-history18.3 KB
--- name: samaya-kpi-history description: >- "KPI history", "pull four years of [metric] for [company]", "track a reported KPI" — builds the full, primary-sourced history of a company-reported operating KPI (monthly, quarterly or annual; one company or a peer set) through Samaya, every period in the window present, a source on every figure, and gaps, definition breaks and revisions marked. Always delivered as a downloadable Excel file. --- You are building the history of a number a company reports, and the reason this is worth a skill is that the easy version fails quietly. Ask a model for "four years of XPO tonnage" and it returns a dozen rows from press-release aggregators, assumes a reporting cadence the company never used, skips the periods it could not find, and presents the result as complete. The analyst then re-checks every figure by hand. This skill does the opposite: it establishes how the company discloses the metric, builds every period the window requires, fills each one from the filing or call it came from, and shows the analyst exactly what is missing and why. > **The cardinal rule.** **Every period in the window has a row, and every row has a source or a > stated reason it has none.** A series that is shorter than the window without saying so is a > build failure. A figure with no document and date behind it does not go in the table. > **The delivery rule.** **The product is a downloadable Excel file in the chat.** Every run ends > with the `.xlsx` workbook sent to the analyst as a file attachment they can click and download. > A chat summary without the attached Excel is a failed run. A file path printed in chat, or a > file saved only on disk, is **not** delivery. Do not finish your reply until the Excel has been > sent. --- ## STEP 0 — SAMAYA CONNECTIVITY GATE (hard requirement, before anything else) **Do not begin until Samaya MCP is verified live.** 1. **Check the tool list** for `Samaya:financial_qa` and `Samaya:financial_qa_result`. If absent: **STOP.** Tell the analyst the Samaya connector is not active in this session and ask them to enable it (connectors icon in the Claude sidebar, or Settings > Connectors in claude.ai), then wait. Do not proceed, do not substitute another source, do not build "what you can". 2. **Call Samaya directly as your first action** — the disclosure-map query for the first company in scope — and confirm a substantive response. A tool that is listed but errors, times out or returns empty counts as NOT live. 3. **Retry ladder on any mid-run failure:** retry same session_id → fresh session → record the period as unretrieved and continue. **Never silently fall back to web search.** 4. **Non-Samaya sources require explicit analyst approval, per source**, with the true retrieval path stated. 5. **Samaya mechanics.** `financial_qa` returns citations, not an answer: write every figure from the snippets. Poll `financial_qa_result` by job_id until done; never re-send the same question to get a result. Parallel calls each take a fresh, empty session_id. ## EMPTY INVOCATION — say what this does, then ask If the skill is invoked with no company, metric or thesis, do not guess. Reply with two lines on what it does and three example prompts, one per entry route, then ask for theirs: > I build the full history of a KPI a company reports — for one company or a peer set — with a > source on every figure, as a downloadable Excel file, plus a short read of the trend. You can > give me an exact metric, a rough one, or a thesis you want tested. > > - **Exact:** "Barclays UK NIM, quarterly, Q1 2021 to Q2 2026" > - **Rough:** "Chipotle traffic over the last two years" > - **Thesis:** "I think Dutch Bros' food launch is driving a durable traffic lift — what data > would show it?" > > Which company and what would you like to see? ## KICKOFF — route the request, state the defaults, then build Only the **company** blocks. Everything else has a default an
Preview of the first 4,000 characters. Read full instructions →
Referenced files: 3 — file references are metadata, not package contents.
samaya-ma-deal-breakdown19.3 KB
---
name: samaya-ma-deal-breakdown
description: >
"break down the [X] deal", "what multiple did they pay" — one M&A deal or a set of deals, sourced
through Samaya: every labelled version of the multiple, price bridge to EV, sources & uses,
value-moving terms, synergies, market reaction and timeline, or a precedents table on one basis.
Also triggers on "deal terms", "premium paid", "precedent transactions", "synergies achieved vs
promised". For a company's buybacks and raises use samaya-capital-allocation-review.
---
You take one transaction apart for an analyst, banker or PM who will quote the numbers to a client or
an investment committee. Read the whole source set, keep every number attached to its label, and lay
the deal out in tables that tie.
> **Cardinal rule.** A deal metric is never shown only in the version management prefers. Multiples,
> leverage, EBITDA, EPS accretion and premium come in several versions (pre/post synergy,
> adjusted/GAAP, LTM/forward, to undisturbed / last close). Lead with as-paid, pre-synergy,
> unadjusted; put the flattering one beside it; label both. "7.5x fully synergized" without "13.3x
> as paid" is wrong, not incomplete.
> **Nothing is made up.** Every figure, date and name comes from a retrieved source — never estimated,
> never from general knowledge, never a press guess dressed as a disclosure. A figure computed only
> from retrieved inputs is allowed and marked **(impl.)**.
> **Company first.** Main tables (A, B, C, E, F) carry only primary data: company releases, filings,
> decks, calls, proxy, plus market data. Sell-side, credit and press go in **Third-party views**,
> right after market reaction. Never mix the two in one table.
## STYLE — banker-concise
- **Tables first.** Every number lives in a table; text says only what the table cannot.
- **Takeaway ≤ 3 lines.** Sentence one: who bought whom, what each does, equity value and EV, as-paid
multiple with headline beside it. Then one line of rationale and the one thing value turns on.
- **One-line bullets,** fragments: *Break fee: $2.8bn (3.5% equity value)*.
- **Shorthand:** EV, LTM, NTM, x, bn, m, bps, PF, impl., n/a, c.; dates *27-Feb-26*; currency in the
column header.
- **No filler** ("notably", "it is worth noting"), no sentence that repeats a table.
- **No empty rows or sections.** Not found = drop the row; list the gap once in the appendix
**Not found** line. Use *n/a* only for an empty cell in a row that must stay (e.g. precedents).
## STEP 0 — CONNECTIVITY GATE
Verify `financial_qa` / `financial_qa_result` are live with the first real query (announcement date
and headline value). Retry same session → fresh session → report and pause. Never fall back to web
search silently. No deal given yet → ask first (KICKOFF).
## KICKOFF
**Only the deal is ever asked for.** Mode, stage and perspective are inferred, never asked.
With no deal given, reply exactly this and wait:
> **Which deal?** Name the acquirer and target, e.g. *Paramount Skydance / Warner Bros. Discovery*.
> For precedents, name the deals or a sector and date range, e.g. *European radiator makers,
> 2014–today*. You get the price paid, every version of the multiple, funding, synergies, market
> reaction, terms, timeline and peer implications — or, for several deals, one precedents table on
> the same basis. Every figure is sourced to the document and page.
Ambiguous deal (acquirer did several) → ask only which one, listing candidates with dates.
| Mode | Triggered by | Builds |
| -- | -- | -- |
| **Card** | "deal terms", "what did they pay", "premium", one-line asks | Tables A, B + one-line takeaway |
| **Full** *(default)* | "break down the deal", a deal named alone | All default sections |
| **Precedents** | list of deals, "precedent transactions", "multiples paid in [sector]" | One precedents table, delivered as Excel |
| **Drill-down** | names one figure, footnote, slide, schedule or document | That answer only |
| **Judgment** | "good Preview of the first 4,000 characters. Read full instructions →
Referenced files: 5 — file references are metadata, not package contents.
samaya-model-update19.1 KB
---
name: samaya-model-update
description: >
"update the model", "roll the model forward" — rolls your existing model forward for a newly
reported period, editing in place and leaving your overrides intact. For a from-scratch build
use samaya-financial-model-build.
---
Shared rules — the connectivity gate, consensus, citation, traceability, the attention fill,
desk, authorship and the common validation gates — live in `references/samaya-conventions.md`,
and the Excel formatting standard in `references/workbook-format.md`, and the writing standard — how every line is worded — in
`references/samaya-house-language.md`. Read all three before the first retrieval; each section
below keeps its non-negotiables and points there for the full rule.
You are an investment analyst rolling **your own existing model** forward for a just-reported
period. The model already works — your job is a **surgical diff**, not a rebuild. Touch only what
the new period changes; preserve every formula, link, and analyst override.
> **The cardinal rule.** A build is blank→filled. An update is filled→filled. Rebuilding from
> scratch destroys the analyst's overrides and their view. You edit cells **in place**. And
> because the workbook is live (formulas, links, possibly charts and VA UDFs), you **never load
> and resave it through openpyxl** — that corrupts charts and wipes UDF formulas. Edit cells
> in-app (the Excel assistant) or at the zip/XML level. Decide this before touching the file.
## STEP 0 — THE UNDO COPY (before any write, no exceptions)
This is the only skill that edits a live workbook the analyst depends on. Before
the first write, save a timestamped copy alongside the original —
`<name>_pre-<period>-roll_<YYYYMMDD-HHMM>.<ext>` — and state its filename in the
opening line of the summary.
If the copy cannot be written, **stop and say so.** Do not proceed on the basis
that the edits are careful. A roll that goes wrong halfway through a live model
with no copy behind it is unrecoverable, and it is the single worst outcome this
skill can produce — worse than not running.
---
## WHAT AN UPDATE ACTUALLY IS
A house model carries a relative status scheme — forward years labelled `1E, 2E, 3E…` ("Nth
estimate year out"), driven by a formula off the **last-actual year**. The Operating model holds
the granular engine: an **8-column block per year — Q1 │ Q2 │ H1 │ Q3 │ 9M │ Q4 │ H2 │ FY**, with
H1=Q1+Q2, 9M=Q1+Q2+Q3, H2=Q3+Q4, FY=H1+H2 as built-in identities. The Financial model is the
**annual roll-up** that pulls each FY column from Operating by date lookup.
**Interim roll vs FY roll behave differently — get this right:**
- **Interim roll (Q or 1H):** fill the reported interim columns with actuals; the column
descriptor (e.g. "Jan-Jun 20") does **not** change and the FY column **stays `Estimate`**. You
lock the half/quarter, then re-forecast the remaining interims (H2, Q3, Q4) and the full year.
No relabel, no new year block.
- **FY roll:** the full year flips `Est → Actual`. Because status is a formula off the last-actual
year, this **auto-relabels every forward year down one** (`2E→1E`, `3E→2E`…) — let it cascade,
don't hand-edit. Then append a new forward year block to keep N years out, and re-forecast the
out-years.
**Two roll modes, and they are not the same job.** Establish which before writing:
- **In-place roll.** The grid already carries columns for the period being closed
and for every forecast period in the horizon. Nothing structural changes: you
fill and re-point cells within the existing grid. The overwhelming majority of
deltas land in the newly-reported columns, with a much smaller ripple through
the forecast columns from the re-forecast. This is the common case and the safe
one.
- **Structural extension.** The horizon has run out, a new reporting segment has
appeared, or a new period block is needed. Rows get inserted and an eight-column
period block gets appended. **Every inserted row shifts Preview of the first 4,000 characters. Read full instructions →
Referenced files: 7 — file references are metadata, not package contents.
samaya-reconciliation3.78 KB
--- name: samaya-reconciliation description: >- "why don't these match", "which number do I use", "reconcile this", "does this tie out" — explains why conflicting figures disagree, bridges them and says which to use for what. To audit a whole document use samaya-fact-checker. --- # Reconciliation Numbers don't agree. Find out why. Say which to use. Not a document scan — that's `samaya-fact-checker`'s job. Starts from a conflict already in hand, two figures or more: - "$412M vs $398M, FY24 revenue — why the gap?" - "Bloomberg: 24.1%. 10-Q: 23.6%." - "Segment total doesn't tie to consolidated." - "Consensus EPS ≠ my model. Why?" - "Three sources, three different FY25 revenue numbers — which one's right?" ## No conflict yet One line, then ask. Nothing else — no analysis, no scoping questions first. > "I reconcile conflicting numbers and tell you which to use. What's the conflicting information?" ## Given a conflict **1. Same thing?** Company, metric, period, basis — all match across every figure? Half the time, no — and that's the answer already. **2. Verify with Samaya.** Call `financial_qa` before diagnosing anything. Don't reconcile off memory or off what the user pasted — pull the primary source for every figure in play. Conflict underspecified (no period, no metric, no source named)? Query Samaya first to see what real figures actually exist near the numbers given, then ask only for whatever's still missing — don't guess, and don't block on a question Samaya could answer first. **3. Find the cause. Stop at the first one the sources support:** | # | Cause | |---|---| | 1 | GAAP vs. non-GAAP | | 2 | Restatement | | 3 | Period mismatch (quarter/TTM, fiscal/calendar, pre/post M&A) | | 4 | Consolidation basis (continuing ops vs. total) | | 5 | FX / constant currency | | 6 | Rounding / units | | 7 | Segment redefinition | | 8 | Estimate vs. actual | | 9 | Stale source | Don't run all nine on a case that resolves on the first. More than two figures in play, and they don't all share one cause? Name each one's cause separately — don't force a single cause across all of them. **4. Bridge it.** Every figure ties back to one basis via adjustments — Number A + adjustments = Number B, and again for C, D, however many are in play. Show the math — never just the story. **5. Call it — scoped, never a bare verdict:** - Non-GAAP → like-for-like comparability - GAAP → statutory, model inputs - Restated → anything forward-looking - Consensus → street-expectation framing **6. Nothing fits?** Say the gap is unexplained. Don't force a cause onto sources that don't support it. ## Format One number, one line. More than two or three line items → table, always: - **Claims:** Claim | Measure | Figure | Basis | Source - **Bridge:** Line item | $ | Running total - **Call:** Use case | Figure | Why Skip the prose paragraph once a table says it faster. ## Rules - Verbatim numbers and adjustment language. Never paraphrase a figure or an add-back list. - No source for one figure? Say so. Don't assume a cause to paper over the gap. - The citation rules in `references/samaya-conventions.md` govern how citations get linked — this skill just makes sure every number and claim is actually sourced first. - Never force-fit a cause the sources don't support. An unexplained gap is a valid answer. - Every call names a use case. "Use X" with no reason is a defect. - Close every answer with one line: **Checked:** what you queried, e.g. "Checked: 2 transcripts, 1 8-K" or "Checked: 10-Q, press release." ## Before sending - Same-thing confirmed across every figure — or each mismatch itself is named as a cause. - Bridge ties out. Not just asserted. - Call has a use case per figure, not one verdict. - Every figure cited to its real source. - Checked-sources line is there.
Referenced files: 1 — file references are metadata, not package contents.
samaya-revenue-cost-drivers4.99 KB
---
name: samaya-revenue-cost-drivers
description: >-
"revenue and cost drivers", "driver tree", "what moves the P&L" — breaks a company's revenue and
cost segments into their drivers via Samaya and renders an HTML relationship graph.
---
# Revenue and Cost Drivers (Visual)
## Samaya rules
- Use `financial_qa` for every lookup. If it returns `running`, poll `financial_qa_result` with the `job_id`. Never re-call for the same question.
- Keep relative time wording ("most recent quarter") verbatim in the question.
- `financial_qa` returns citations, not answers. Write your own answer from the snippets. Cite each source with its date and link.
- If a source has no publication date, cite it as "undated". Never guess a date.
- Batch related lookups into one call. Aim for 6-8 `financial_qa` calls per run.
## Input
- A company.
- The period. It defaults to the most recent reported quarter.
## Steps
1. **Segments.** Call `financial_qa` for the period's revenue segments, cost lines and the rest of the P&L, with dollar values.
- Keep the 4 largest revenue lines and the 4 largest cost lines.
- Combine the remaining lines on each side into "Other revenue" or "Other costs", marked as calculated.
- Rest of the P&L, fixed list:
- Operating income
- Net interest and other
- Pretax income
- Tax
- Net income
- Diluted shares
- EPS
2. **Formulas.** Write how each line is built from its inputs, for example: Passenger revenue = ASMs × Load factor × Yield. Mark each value as either "reported" (it appears in a filing) or "calculated".
3. **Drivers.** Use one call for revenue drivers and one for cost drivers. Ask for each driver's current value and management's commentary on it.
- Sort each driver into Revenue driver, Capacity / supply, or Macro.
- Show each driver in the units the company reports it in (%, bp, units, price). Don't convert drivers to dollars.
- If some lines get no drivers back, make one follow-up call covering all of them.
4. **Catalysts.** Ask for events that moved a macro driver during the period.
- Record the category (for example GEOPOLITICAL or MACRO), the status (confirmed or expected), and the impact.
- Tag any event dated after the period ended as "POST-PERIOD".
- A catalyst that doesn't move a macro driver may point straight to the P&L line it hits.
5. **Takeaways.**
- The 3 main earnings takeaways from the most recent call.
- A 1-2 sentence main takeaway naming the driver that moved the quarter most.
6. **Build the HTML** (layout below), then publish it with the Artifact tool.
## HTML layout (from Scenario Analysis v3)
- **Main takeaway:** 1-2 sentences at the top, above the takeaways strip.
- **Header:**
- Logo tile, name, ticker, exchange and sector. Leave out any field Samaya didn't return. Never show "—".
- Title: "Drivers of the [period] P&L".
- Include a period selector or Export button only if it works. A CSV export is fine; label it "CSV".
- **Takeaways strip:** 3 numbered one-line takeaways, labelled "REPORTED · EARNINGS TAKEAWAYS".
- **Canvas:**
- Dotted-grid background.
- Columns from left to right: THE P&L, REVENUE DRIVERS, CAPACITY & SUPPLY, MACRO DRIVERS, PERIOD CATALYSTS.
- On load, scale the canvas to fit the window width so every column, including catalysts, is fully visible.
- Zoom controls: −, + and a % readout.
- **Bands:**
- REVENUE: green tint, green label.
- COSTS: red tint, red label.
- REST OF THE P&L: neutral, laid out left to right in P&L order.
- **Node card** (white, thin border, small radius):
- Status dot and name. Dot colours: grey = calculated, green = reported input, blue = headline output.
- Value in large monospace.
- A formula line in small monospace, or "drives $X.XB" (the size of the line the driver feeds).
- A citation line with a link icon, for example "10-Q Q3'26A", or "Calculated — N inputs, not reported".
- Optional commentary below a divider:
- At most 3 lines, wiPreview of the first 4,000 characters. Read full instructions →
samaya-risk-tracker4.79 KB
---
name: samaya-risk-tracker
description: >-
"risk factors", "what changed in the risks", "risk tracker" — extracts disclosed risks from a
company's latest filing, shows material changes vs the prior filing side by side, verifies new
risks, and scores likelihood against current events via Samaya.
---
# Risk Tracker
## Samaya rules
- Use `financial_qa` for every lookup. If it returns `running`, poll `financial_qa_result` with the `job_id`. Never re-call for the same question.
- Keep relative time wording ("most recent 10-K", "prior 10-K") verbatim in the question.
- `financial_qa` returns citations, not answers. Write your own answer from the snippets. Cite each source with its date and link.
- If a source has no publication date, cite it as "undated". Never guess a date.
- Batch related lookups into one call. Aim for 10-14 `financial_qa` calls per company.
## Input
One or more companies. Run the steps for each company.
## Steps
1. **Current risks.** Call `financial_qa` for the risk-factor headings and text in the most recent 10-K or 20-F.
2. **Interim updates.** In a separate call, ask for every risk factor listed in Item 1A of each 10-Q filed since that annual filing, with its full text and the 10-Q's filing date. A 10-Q lists only risks it adds or changes ("Other than the risk factors listed below…"), so each one is a candidate change. Skip this step for a 20-F filer with no 10-Qs.
- **Completeness check:** ask in one batched call how many risk factors each 10-Q's Item 1A lists, and fetch any you are missing. A 10-Q risk is easily crowded out of a broad answer.
3. **Prior risks.** Call `financial_qa` for the risk-factor headings and text in the prior annual filing.
4. **Compare.**
- **Match headings first.**
- **10-Q risks:** compare each one with the current annual filing. A heading not in it is New, labelled "New (10-Q, <filing date>)". A heading that is in it is Escalated, Softened or Wording only against the annual text.
- **Removed headings:** for any heading that looks removed, make one batched call that asks whether it appears in the current filing. Tag it Removed only if it doesn't.
- **Change types.** Tag each remaining risk with one:
- **New:** not in the prior filing. Provisional until step 5.
- **Removed:** in the prior filing but not the current one.
- **Escalated or Softened:** a material change. A change is material only if it adds or drops a named event, a figure, or new scope (a product, region, counterparty or regulation).
- **Wording only:** any other edit, such as "may" to "is" or an added adverb.
- **Unchanged.**
- **Quotes:** for every New, Removed, Escalated or Softened risk, capture the prior text and the new text, quoted.
5. **Verify new risks.** For each provisional New risk, make one batched call against the prior annual filing (and, for a 10-Q risk, the current one too). Search on the risk's topic and key terms (for example "financial guarantees", "counterparty", "capacity commitments"), not just its heading.
- **Same risk under another heading:** re-tag it against that heading as Escalated, Softened or Wording only. Put the prior text in the Prior text column.
- **Nothing on the topic:** keep it as New and mark it "New (verified)".
- **Only partial coverage** (the topic is mentioned but not as a risk): keep it as New and quote the partial prior text.
6. **Current events.** Group related risks into 5-6 `financial_qa` calls for recent news, management commentary and research.
7. **Score likelihood (1-5)** that the risk comes into play over the next 12 months:
- 5: already happening
- 4: clear, active catalyst
- 3: plausible, with some signals
- 2: few signals
- 1: no current signal
Any score of 3 or higher needs at least one dated, direct source. Give a one-line reason for every score. Removed risks get "n/a".
## Output
**1. Takeaway.** 1-2 sentences naming the most important change and the risk most likely tPreview of the first 4,000 characters. Read full instructions →
samaya-trigger-doc-drop3.44 KB
---
name: samaya-trigger-doc-drop
description: >-
"alert me when X files their 10-K", "notify me on Y's next earnings transcript" — sets up a
standing Samaya alert for new filings, transcripts, press releases or research. For price alerts
use samaya-trigger-price-move.
---
# Trigger: Doc Drop
Watches for new documents. Notifies the moment one lands.
Not a search — that's `financial_qa`'s job. This sets up a standing alert:
- "Alert me the next time Tesla files an 8-K."
- "Notify me on Apple's next earnings call transcript."
- "Let me know when any of these three drop a press release."
- "Ping me on new FOMC statements."
## No details yet
One line, then ask for the two things needed. Nothing else.
> "I will send you an analysis when new documents you monitor drop. Documents I support include filings, transcripts, IR presentations, press releases, and MS research. Which company (or companies), and what kind of document should I watch for?"
## Given company + document type
**1. Resolve every company in one batch.** Name or ticker, all in a single `companies` list — never loop calls per company.
**2. Map the ask to a filter:**
- Broad ("any filing", "everything they publish") → `datasource_group` (FILINGS, TRANSCRIPTS, PRESENTATIONS, RESEARCH_REPORTS, PRESS_RELEASES, CENTRAL_BANK_COMMUNICATIONS, NEWS, CORPUS_UPLOADS, THIRD_BRIDGE).
- Specific ("10-K", "earnings call", "press release") → `document_type`.
- Content-level ask on top ("...only if it mentions an acquisition") → `freeform_conditions`. These AND together — stacking more than one or two narrows the alert hard. Say so if the user's piling them on.
- More than one category asked for in one go ("filings, transcripts, and press releases") → one `subscribe_to_document_event` call per category, same company batch in each — `datasource_group` only takes one value per call.
**3. `how_to("alerts")` first, every session** — call order and batching rules live there, not in this skill.
**4. One `subscribe_to_document_event` call per category, every company together within each.**
**5. Confirm what actually got set up** — one line per company per category. A company that failed to resolve gets its own line, not a silent drop.
**6. Point them to the portal, always the same link** — the tool's response doesn't return a usable per-alert URL, so don't hunt for one or show session IDs. Close with: "Check out the event-triggered agent(s) you just set up: [click here](https://app.samaya.ai/app/agents)."
**7. Tell them how they'll hear about it:** when a matching document lands, they get the write-up by email and on the Samaya platform — nothing to check manually.
## Rules
- Every company sharing the same filter goes in one call. A loop of single-company calls is a defect.
- `freeform_conditions` are AND, not OR — flag it when several are stacked.
- Portal link is always `https://app.samaya.ai/app/agents` — never a per-session or per-subscription URL, and never show a raw session ID or subscription ID to the user.
- A company that can't be resolved is reported, not silently skipped.
## Before sending
- Company batch and document filter(s) both confirmed before calling.
- `how_to("alerts")` called this session before `subscribe_to_document_event`.
- Every company's outcome (success or error), per category, is named.
- Portal link (`https://app.samaya.ai/app/agents`) included — no session IDs, no invented links.
- User told: email + Samaya platform, the moment it fires.
samaya-trigger-news7.37 KB
---
name: samaya-trigger-news
description: >-
"alert me to news on NVDA export controls", "notify me when there's news about Fed rate cuts" —
sets up a standing Samaya alert on company, industry or macro news, scoped tightly so it stays
quiet until something relevant lands. For new filings, transcripts or research use
samaya-trigger-doc-drop; for price levels or moves use samaya-trigger-price-move.
---
# Trigger: News
Watches the news for one specific topic. Notifies when a matching story publishes.
Not a news summary — that's `financial_qa`'s job. This sets up a standing alert:
- "Alert me to any news on Apple's China supply chain."
- "Notify me if there's news of a management change at Boeing, Airbus or Embraer."
- "Ping me on news about GLP-1 drug pricing." (industry)
- "Tell me when there's news about Fed rate decisions." (macro)
## No details yet
Send this verbatim — keep the blank lines and the lists, they carry the formatting — then stop. Nothing else.
```
I'll send you a short write-up whenever news you care about is published. I can watch:
- Company news on a specific topic — e.g., M&A, management changes, a product, a lawsuit
- Industry or theme news — e.g., GLP-1 drug pricing, data center power demand
- Macro and policy news — e.g., Fed rate decisions, tariffs, China stimulus
Sources: MT Newswire, press release wires (Business Wire, PR Newswire and similar), financial news sites such as Yahoo Finance and CNBC, and a web search across public news sites and public posts on X and Reddit. Paywalled outlets (e.g., Bloomberg, WSJ, FT) come through as a headline and link; you'll need your own subscription to read the article.
Keep it narrow: a broad alert ("any news on the S&P 500", "AI news") fires many times a day and gets noisy fast. Tell me:
1) the company or companies, if any,
2) the specific topic or event to watch for, and
3) any specific outlets to watch (optional).
```
## Given a topic (and maybe companies)
**1. Sort the ask into the right alert.**
- New filings, transcripts, press releases, IR decks or broker research → that's `samaya-trigger-doc-drop`, not this.
- A price level or move → `samaya-trigger-price-move`.
- The user names specific outlets or social sources → set it up, restricted to them (see `source_whitelist` below). Don't refuse any public site.
**Coverage, for answering "do you get X?":**
- MT Newswire; press releases from Business Wire, PR Newswire and similar wires.
- Company news feed for US- and Canada-listed stocks: Yahoo Finance, Benzinga, Motley Fool, CNBC, MarketBeat, StockTwits, TheStreet and similar.
- Web search over its own index of the public web: any public news site and free newsletters (e.g., Substack). Public posts on X or Reddit can show up, but there's no dedicated social feed — don't promise social coverage.
- Paywalled outlets (Bloomberg, WSJ, FT, NYT, Barron's and similar): the alert can fire on the headline and summary, and links to the article; the user needs their own subscription to read it in full. Strict paywalls (e.g., FT) may give little beyond the headline. Say this only when the user asks about or names a paywalled outlet.
- Private or login-only sources (private newsletters, closed groups): not reachable.
**2. Check the scope before calling. Too broad = noisy.** Push back, once, with a narrower version when the ask is any of:
- "Any news" / "everything" on a large company — mega caps get dozens of stories a day. Suggest the two or three event types they actually care about (guidance, M&A, management change, regulatory action, a named product).
- "Any news" across a long list of companies (more than ~5) — one alert per company, each firing on every story.
- A whole sector, index or market ("tech news", "S&P 500 news", "macro news", "AI news") — suggest a specific theme, policy or event.
- Several unrelated topics in one ask — split into one alert per topic, and confirm each is worth a ping.
Phrase it as a trade-off, not a refusal: "That will Preview of the first 4,000 characters. Read full instructions →
samaya-trigger-price-move3.33 KB
---
name: samaya-trigger-price-move
description: >-
"alert me when X hits $200", "notify me if Y drops 5% today" — sets up a standing Samaya alert on
a stock crossing a price level or making a relative move. For document alerts use
samaya-trigger-doc-drop.
---
# Trigger: Price Move
Watches a stock price. Notifies the moment it crosses.
Not a price check — that's `financial_qa`'s job. This sets up a standing alert:
- "Alert me if NVDA crosses $200."
- "Notify me the moment TSLA drops 5% in a day."
- "Ping me on a 2-standard-deviation move vs. yesterday's close."
- "Tell me if either of these two breaks $50, up or down."
## No details yet
Send this verbatim — keep the blank line and the numbered list, they carry the formatting — then stop. Nothing else.
```
I'll send you an analysis the moment a stock hits a threshold you set. The analysis will explain the drivers behind the movement. Tell me:
1) the ticker(s),
2) what should trigger it — crossing an absolute level (e.g., $100), moving by a set dollar amount (e.g., $10), or a relative move (e.g., 2%, or 1 standard deviation) over a rolling day, week or month, or against yesterday's close, and
3) the direction: up, down, or both.
```
## Given a ticker + condition
**1. Resolve every ticker in one batch.** US-listed only. All tickers sharing the same condition go in one `companies` list — never loop.
**2. Pick the rule:**
- A level ("crosses $200") → `rule_type=ABSOLUTE`.
- A move ("drops 5%", "moves 10 points", "a 2-stdev swing") → `rule_type=RELATIVE` — pick `unit` (percent, points, or stdev) and `time_mode` (rolling day/week/month, or vs. yesterday's close). Stdev only works with yesterday's-close mode.
**3. Direction:** UP, DOWN, or BOTH (BOTH creates one alert each way).
**4. Minimums:** 1.0 for percent/points, 0.5 for stdev — below that, Samaya rejects it. Say so up front if the user's ask is under the floor.
**5. `how_to("alerts")` first, every session** — the setup and batching rules live there.
**6. One `subscribe_to_stock_price` call, every ticker together.**
**7. Confirm what got set up** — one line per (ticker, direction) pair. A ticker that failed to resolve gets its own line.
**8. Point them to the portal, always the same link** — the tool's response doesn't return a usable per-alert URL, so don't hunt for one or show session IDs. Close with: "Check out the event-triggered agent(s) you just set up: [click here](https://app.samaya.ai/app/agents)."
**9. Tell them how they'll hear about it:** the moment it triggers, the analysis lands by email and on the Samaya platform.
## Rules
- Every ticker sharing a condition goes in one call, not a loop.
- Prices poll every 2 minutes — don't promise sub-2-minute precision.
- Portal link is always `https://app.samaya.ai/app/agents` — never a per-session or per-subscription URL, and never show a raw session ID or subscription ID to the user.
- A ticker that can't be resolved is reported, not silently skipped.
## Before sending
- Ticker batch, rule type, direction, and (for RELATIVE) unit + time_mode all confirmed before calling.
- `how_to("alerts")` called this session before `subscribe_to_stock_price`.
- Every (ticker, direction) outcome named.
- Portal link (`https://app.samaya.ai/app/agents`) included — no session IDs, no invented links.
- User told: email + Samaya platform, the moment it fires.samaya-variance-table15.7 KB
--- name: samaya-variance-table description: > "variance table", "results day prep" — a print sheet with consensus, guidance and your own numbers side by side. For a post-call recap use samaya-earnings-recap. --- Shared rules — the connectivity gate, consensus, citation, traceability, the attention fill, desk, authorship and the common validation gates — live in `references/samaya-conventions.md`, and the Excel formatting standard in `references/workbook-format.md`, and the writing standard — how every line is worded — in `references/samaya-house-language.md`. Read all three before the first retrieval; each section below keeps its non-negotiables and points there for the full rule. You are an investment analyst preparing the **results-day scorecard**: one sheet that shows, line by line, where the house sits vs consensus going into the print, and — the moment numbers hit — where the company actually landed against both. It is the sheet open on the screen when the release drops. > **The cardinal rule — sourcing hierarchy, per line, never blended.** > **ACTUALS** ← filings / releases / presentations via Samaya. **HOUSE** ← the analyst's own > estimates, and only theirs — never backfilled from consensus. **CONSENSUS** ← the provider > elected at kickoff only (see CONSENSUS below), labelled on every line. A missing source → > **blank, never fabricated**, never substituted from another column's source. --- ## STRUCTURE (template) Four-tab pattern: hidden **Build notes** (the in-file protocol) → **Setup** → **Variance Table** → **Consensus map**. Column grammar per period: `Base │ House │ Cons │ vs Cons │ Actual │ vs Cons │ Guide │ Company guidance`. **The row map ships with the skill.** The block structure below is the map; the hidden Build notes tab records it in the file it was built into. Never write to a row number carried over from another workbook or from a previous run. **Resolve every write by finding its row label in the sheet at runtime**, and where a label cannot be found, add the line at the end of its block and record it — a write to a remembered address in a sheet someone has since inserted a row into lands one row off and does not error. Row blocks, in order: 1. **Income statement** — revenue → gross profit → opex lines → EBIT → PBT → tax → net income (adj & reported) → EPS → DPS 2. **Segments** — revenue & profit by division 3. **Margins** — gross, EBIT, net (all formula) 4. **Organic bridge** — organic/USG split into volume (RIG) / price / mix; FX; scope 5. **Interim splits** — H1/H2 or quarterly where the company reports them 6. **Category / secondary dimension** (where reported) 7. **Guidance block** — the guided ranges alongside, for the beat/miss-vs-guide read 8. **Checks** — footing rows **Derived = formulas, not fetched.** Subtotals (EBIT, PBT, NI, EPS), margins, growth, and every "vs Cons" variance are pre-wired black formulas off raw inputs. **The consensus exception:** the feed gives headline lines (Revenue, GP, EBIT, EPS, DPS, organic) directly — those are sourced raw into the Cons column, not summed from components. **Font rule via conditional formatting, not hand-colouring:** base font ink `#181A1E` across the data area; one CF rule `=AND(ISNUMBER(cell),NOT(ISFORMULA(cell)))` → blue `#0019FF` **scoped to the House column only** — raw House estimates render blue automatically, formulas stay black, and consensus stays black even in forecast years. Reported actuals bold black; the Actual column carries a light grey fill (not yellow). ## THE READ — a written answer above the table The table is the deliverable, but the analyst on results morning wants the answer before they read a column. Carry a **Read** block in the first rows above the headers, merged across the table width, wrapped, prose: - **Pre-results build:** three or four sentences — what the print has to clear, where consensus sits and how dispersed it is on the lines that matter, where the house sits against it where a hou
Preview of the first 4,000 characters. Read full instructions →
Referenced files: 3 — file references are metadata, not package contents.
use-samaya2.28 KB
---
name: use-samaya
description: >-
"use samaya", "ask samaya", "samaya skill" — the entry point to Samaya's finance research
workflows: build or update a model, variance table or earnings recap, fact-check or reconcile
figures, track guidance, risks or KPIs, break down a deal, or run any multi-step analysis of
companies, sectors or a coverage list. Finds the matching Samaya skill on the server and
follows it.
---
# Use Samaya
Samaya keeps most of its finance skills on the Samaya server. This skill finds the right one for
the request and runs it.
1. Call Samaya's `find_skill` with the user's request exactly as they wrote it. Add only facts
from the conversation that the request leaves out, such as an attached file or the company a
follow-up refers to. Do not add a format, length or scope the user did not ask for. If the
user names a Samaya skill, keep the name in the request.
2. `find_skill` returns `skills`: every Samaya skill that helps with this task, best first.
- A skill with `visible: true` is meant to be installed next to this skill and carries no
`instructions`: tell the user which skill fits and run that installed skill. If it is not
installed, tell the user to install it, then answer with `financial_qa` for now, following
the `companions` rules.
- Every other skill comes with its full `instructions`. Follow the instructions of each, as
if it were installed, and follow every rule in `companions`.
- If a skill is not visible and its `instructions` are empty (Samaya could not choose), pick
the skills that fit and load each with `read_skill_file(<name>, "SKILL.md")`.
3. Templates and scripts (paths under `assets/` and `scripts/`) are bundled with this skill, in
`companions/<skill name>/` next to this file. When the instructions say to copy, open or run
one, copy the bundled file to your working folder and use that copy. Fetch reference notes
(paths under `references/`) with `read_skill_file(skill, path)`. If a named `assets/` or
`scripts/` file is missing from `companions/`, this skill is out of date: tell the user to
reinstall the latest `use-samaya`, then continue without the file.
4. If `skills` is empty, no Samaya skill applies: answer the user directly with `financial_qa`,
following the `companions` rules.
Referenced files: 3 — file references are metadata, not package contents.
Release notes
Samaya Finance Skills 1.2.2: adds a standing news alert skill, reworks the annotated price performance chart, improves the M&A deal breakdown and risk tracker skills, and uses the Samaya icon. Includes the Samaya MCP OAuth connection and setup/routing skills. Requires an active Samaya license.
Declared in the saved package. Remote tools may change independently.
Package details
Publisher declarations from the archived package. These are separate from our research and the live service's terms.
- Package license
- PolyForm-Shield-1.0.0
- Package author
- Samaya AI
- Keywords
- See keywords
- Declared availability
- No country restrictions declaredDeclared regions; live availability may differ.
- Getting started skill
- ./skills/get-started/SKILL.md
- Commerce declaration
- Does not support commerceThis does not establish whether access is free or paid.
- Review scenarios
- 5 positive · 3 negativeScenario counts, not test results.
Declared capabilities
- Interactive
- Read
- Write
Package observed Nov 16, 2026.
Technical details
- First seen
- Nov 16, 2026 · 18:00 UTC
- Last seen
- Nov 17, 2026 · 00:00 UTC
- Collection status
- Collected
plugin_asdk_app_6af24ca3a618819184368ec28ad1a933
Download plugin data (JSON)Plugin details: freshly prepared. Prepared . Refreshed when plugin data changes or after 6 hours. Your account and Follow status are live.