Grain

View saved version →

---
name: follow-up
description: Draft sendable work from Grain meetings, including follow-up emails, client recaps, internal updates, briefs, decision logs, evidence packs, and tickets. Use when the user wants to send, post, or file something grounded in one or more meetings.
---

# Follow Up

Turn meetings into work the user can send or file without removing AI residue.

Read:

- `references/artifacts.md` for audience, drafting, and delivery standards.
- `references/retrieval.md` when the source meetings are not already identified.
- `references/evidence.md` for citations and claim discipline.

## Workflow

1. Identify every source meeting. For "the call this morning," list the user's attended meetings for today. For an account or topic, apply the retrieval playbook. Read notes and action items before transcripts.
2. Infer the audience and artifact type. State one brief assumption when needed and draft before asking follow-up questions.
3. Draft freely from the evidence, then verify every commitment, owner, date, and material claim against structured action items or the transcript.
4. Deliver only the clean artifact and any compact source line required by the audience.

For external audiences, omit internal candor and Grain links unless the user confirms the links are shareable. For multi-meeting artifacts, state the coverage. Never send, post, create a ticket, or write to another system without explicit confirmation.

Referenced files: 4 — file references are metadata, not package contents.

View saved version →

---
name: heads-up
description: Quietly check Grain meeting history when the user's current work names a specific customer, person, company, account, or project, and surface only evidence that would materially change the work. Use for customer-facing drafts, account decisions, or claims about what someone wanted, said, or agreed to. Skip personal and social messages.
---

# Heads Up

Use Grain as ambient organizational memory without becoming noisy.

Read `references/evidence.md` before surfacing anything.

## Rules

1. Check cheaply and silently. Resolve the named entity, then inspect recent notes or run one targeted transcript search. Do not run a deep sweep for this side-check.
2. If Grain is unavailable or unauthorized, stop silently. Never interrupt unrelated work with an authentication prompt.
3. Surface a heads-up only when the record would change the user's action because they are about to contradict a commitment, overlook a directly stated position, relitigate an existing decision, or use a conflicting date, number, or claim.
4. Write one to three cited sentences, set them apart from the main work, and then continue the user's task.
5. If nothing material appears, say nothing about Grain or the check.
6. Never create, tag, send, or update anything through this skill.

Referenced files: 2 — file references are metadata, not package contents.

View saved version →

---
name: prep-me
description: Prepare the user for an upcoming conversation using Grain meeting history. Use when the user asks for meeting prep or a concise briefing about a meeting, person, company, account, or project before a conversation.
---

# Prep Me

Orient the user for the next conversation without burying them. Keep the brief to one screen when possible.

Read:

- `references/retrieval.md` for entity resolution and account-first scoping.
- `references/evidence.md` for fact, interpretation, and citation rules.

## Workflow

1. Resolve the counterpart and account. If the target remains unclear, ask for the meeting title or participants. Do not require calendar access.
2. Recover the current state, what each side owes, open questions, sensitivities, and anything the counterpart said that deserves acknowledgment. Start account prep with the company dossier when available.
3. Verify commitments against structured action items. Mark "worth raising" and "careful with" items as interpretation rather than meeting fact.

## Output

```text
Prep: [meeting, person, or account]

Where it stands:
...

Open between you:
- You owe them: ...
- They owe you: ...

Worth raising:
- ...

Careful with:
- ...

Sources: [meeting, date, link]
```

Referenced files: 3 — file references are metadata, not package contents.

View saved version →

---
name: voice-of-customer
description: Synthesize patterns across multiple customers using Grain meetings and, when connected, support conversations or customer email. Use for questions about what customers are saying, emerging workflows, needs, feature requests, confusion, bugs, churn signals, topic analysis, or a weekly customer-insights digest. Use where-things-stand instead for one account.
---

# Voice of Customer

Turn customer conversations into a defensible cross-customer evidence report. Grain is the primary source; connected support and email sources are optional enrichment.

Read:

- `references/customer-research.md` for source collection, deduplication, categories, trends, and report structure.
- `references/retrieval.md` for Grain retrieval.
- `references/evidence.md` for quotations, denominators, confidence, and coverage honesty.

## Workflow

1. Select the mode: topic analysis, signal scan, or weekly digest. Resolve the time window and state its exact dates.
2. Inventory available sources. Query Grain. Use support or email only when connected; do not block on unavailable optional sources.
3. Collect targeted customer-facing evidence and build the evidence ledger. Exclude internal discussion unless it contains the customer's own words.
4. Deduplicate by customer and concept before counting. Spot-check the three highest-frequency findings against source IDs.
5. Synthesize the findings with unique-customer denominators, exact quotes when available, `Explicit` or `Inferred` labels, source links, and disclosed coverage gaps.
6. Compare to a previous report only when one is accessible. Otherwise mark the report as a baseline.
7. Return the report in chat. Save, post, message, or schedule it only after explicit confirmation.

Do not turn one account into a market pattern. Do not treat repeated mentions by one customer as several customers. Keep evidence separate from interpretation and recommendations.

Referenced files: 4 — file references are metadata, not package contents.

View saved version →

---
name: what-did-we-say
description: Answer factual questions from the Grain meeting record with cited evidence. Use when the user asks what was said, decided, promised, or why; asks whether someone actually agreed to a claim; or asks about a theme within specified meetings, an account, or a project. Use voice-of-customer instead for structured patterns across many customers.
---

# What Did We Say

Answer from the meeting record with evidence the user can open.

Read:

- `references/retrieval.md` for entity resolution, counter-evidence searches, and efficient retrieval.
- `references/evidence.md` for citations, verdicts, denominators, and honest gaps.

## Modes

### Lookup

Give a short answer followed by cited evidence. Add a separate `My read` section only when interpretation is necessary.

### Verification

Give every claim a verdict: `Supported`, `Contradicted`, or `Not found`. Show the exact or clearly marked paraphrased evidence for each verdict.

### Scoped patterns

For patterns within named meetings, an account, or a project, state the denominator and signal strength. Route broad product-wide or cross-customer research to `voice-of-customer`.

Do not invent an answer when the record is incomplete. State the scope searched and the most promising next search.

Referenced files: 3 — file references are metadata, not package contents.

View saved version →

---
name: what-do-i-owe
description: Recover open commitments from Grain meetings as a ledger of what the user owes, what others owe them, and unowned loose ends. Use for action-item sweeps, promises, weekly reviews, returns from time off, and preparation for a one-on-one.
---

# What Do I Owe

Aggregate commitments across meetings, deduplicate them, separate mine from theirs, and flag likely staleness.

Read:

- `references/retrieval.md` for scoping and source order.
- `references/evidence.md` for confidence, staleness, and source requirements.

## Workflow

1. Resolve the user's identity. Default to 7 days, or 30 days for a return from time off or an "everything open" request. State the window.
2. Pull structured meeting action items first, then notes, then targeted transcript searches for commitment language.
3. Deduplicate the same commitment across extraction layers. Group items into owed by the user, owed to the user, and unowned loose ends.
4. Retain only items tied to a specific action item or utterance. Flag old items as potentially stale.

## Output

```text
Open commitments ([window], [scope])

You owe:
- [action] — [due date] (source)

Owed to you:
- [person]: [action] — (source)

Loose ends:
- [action] — (source)
```

Never create or update tasks in another system without explicit confirmation.

Referenced files: 3 — file references are metadata, not package contents.

View saved version →

---
name: where-things-stand
description: Reconstruct current state from Grain meetings over time. Use when the user asks where things stand, what changed, what remains open, or what the latest is for one account, project, topic, or person; when they need a catch-up after time away; or when they ask about recent recorded meetings.
---

# Where Things Stand

Fold conversations over time into current state, changes, and open work. Do not produce a meeting-by-meeting recap.

Read:

- `references/retrieval.md` for scope defaults, inexpensive retrieval, and latency control.
- `references/evidence.md` for supersession, conflicts, and coverage honesty.

## Workflow

1. Scope by account or person when possible and state the time window.
2. Read notes chronologically with the newest last. Use action items for open work and transcript search only for specific gaps.
3. Distinguish decided, discussed, and mentioned. Note reversals and conflicting meetings.
4. Complete the answer when each open item has an owner and source.

## Output

```text
Where things stand: [scope and window]

Current state:
- ...

What changed:
- ... (meeting, date, link)

Open:
- ... (owner, source)

Worth opening:
- [meeting, date, link] — [reason]
```

Keep it readable in under two minutes. Drop empty sections. "Nothing significant changed" is a complete answer when supported.

Referenced files: 3 — file references are metadata, not package contents.