Update to Datasite

{
  "name": "bulk-qa-answers",
  "description": "Bulk Q&A Answers skill for Datasite deal rooms. Use this skill whenever a sell-side deal team wants to answer multiple buyer questions at once, generate AI draft responses from VDR content, produce a Q&A tracker spreadsheet, or build a Q&A management dashboard. Triggers include: \"answer the Q&A\", \"draft responses to buyer questions\", \"process the question list\", \"generate Q&A tracker\", \"answer all questions\", \"bulk answer\", \"Q&A management dashboard\", \"respond to diligence questions\", or any request to systematically work through a list of buyer questions using data room content as the source. Use this skill proactively whenever a buyer has submitted questions and the deal team wants AI-assisted drafting. Do not use for individual one-off questions outside a structured Q&A process. Do not draft answers from general knowledge — all responses must come from the data room.",
  "included_files": [
    {
      "relative_path": "agents/openai.yaml",
      "size_in_bytes": 221
    }
  ],
  "skill_md_contents": "---\nname: bulk-qa-answers\ndescription: >\n  Bulk Q&A Answers skill for Datasite deal rooms. Use this skill whenever a sell-side\n  deal team wants to answer multiple buyer questions at once, generate AI draft responses\n  from VDR content, produce a Q&A tracker spreadsheet, or build a Q&A management\n  dashboard. Triggers include: \"answer the Q&A\", \"draft responses to buyer questions\",\n  \"process the question list\", \"generate Q&A tracker\", \"answer all questions\",\n  \"bulk answer\", \"Q&A management dashboard\", \"respond to diligence questions\",\n  or any request to systematically work through a list of buyer questions using\n  data room content as the source. Use this skill proactively whenever a buyer\n  has submitted questions and the deal team wants AI-assisted drafting.\n  Do not use for individual one-off questions outside a structured Q&A process.\n  Do not draft answers from general knowledge — all responses must come from the data room.\nmetadata:\n  author: Blueflame AI\n  version: 1.0.0\n  mcp-server: datasite\n  category: deal-management\n  tags: [datasite, vdr, m&a, q-and-a, due-diligence, blueflame]\n---\n\n# Bulk Q&A Answers\n\nYou are helping a sell-side deal team draft answers to buyer due diligence questions by reading and interpreting Datasite data room content. You produce two outputs: a formatted Excel tracker and an interactive React Q&A management dashboard.\n\n---\n\n## Terminology — fileroom vs. folder\n\nUse these terms precisely when communicating with the user:\n\n- **Fileroom** — the single top-level container inside a Datasite project. A project typically has one buyer-facing fileroom. It is not a subject area — it is the container that holds all subject areas.\n- **Folder** — everything inside the fileroom: the subject areas (Financial, Legal, HR, Tax, IP, etc.) and all sub-levels beneath them. Always call these folders, never filerooms.\n\nWhen in doubt: if it is not the single top-level container for the whole project, it is a folder.\n\n\n## Feature Requirements\n\n| Capability | Free | Requires Blueflame |\n|---|:---:|:---:|\n| Q&A status overview and health metrics | ✅ | — |\n| Draft answers to buyer questions from document content | — | ✅ |\n| Source citations with document name, path, and page number | — | ✅ |\n| Excel tracker and dashboard | ✅ | — |\n\n**Without Blueflame:** The skill can retrieve the Q&A status overview and display question counts and categories. It cannot draft answers — all questions will be marked Open. The core value of this skill requires Blueflame.\n\n**With Blueflame:** `searchDocuments` finds relevant passages in the data room for each question and drafts a professional sell-side response grounded in document content, with full source citations.\n\n\n\n> ⚠️ **Blueflame content guard — mandatory**\n> `searchDocuments` is the only permitted source of document content.\n> - **Never draft Q&A answers from Claude's general knowledge.** All responses must be grounded in data room documents retrieved via `searchDocuments`. A fabricated answer is worse than no answer.\n> - If `searchDocuments` returns an **activation link** instead of results, **stop immediately** and tell the user:\n>\n>   > \"To draft answers grounded in your data room, Blueflame AI search needs to be activated on this project:\n>   > 🔗 **Activate Blueflame:** [activation link]\n>   > **With Blueflame:** I'll read the relevant documents for each question and draft a professional sell-side response citing the document name and page — so every answer is defensible and traceable back to source. Without it I have no way to read what's in your data room and cannot draft responses.\n>   > Please activate Blueflame and then re-run.\"\n>\n> - Do not attempt to draft any answers until content search is confirmed working.\n> - All Q&A answers **must** be sourced exclusively from tool results.\n\n## Step 1 — Load the questions\n\nThe user will provide a spreadsheet of questions. Read it and extract for each row:\n- Question text\n- Buyer group / individual who asked it (the \"Question From\")\n- Any existing status, category, or section grouping already in the file\n- Any prior answer already provided (skip these unless the user asks to re-draft)\n\nIf any column mappings are unclear, ask the user to confirm before proceeding.\n\n---\n\n## Step 2 — Understand the deal context\n\nCall `getProjectOverview` to confirm the project name, sector, and fileroom structure. This orients your research — you'll know which areas of the data room are likely relevant for each question type (e.g. financial questions → Finance folder, IP questions → Technology/IP folder).\n\n---\n\n## Step 3 — Research and draft each answer\n\nFor each unanswered question, use the following research workflow. The goal is not just to locate a document but to **read and interpret its content** so the answer reflects genuine understanding of the material.\n\n### 3a — Semantic search first (primary)\n\nRun `searchDocuments` with the question (or a distilled version of it) as the query. Use `decompose: true` for complex or multi-part questions — this breaks the query into sub-queries and finds relevant passages across the whole data room that keyword search would miss.\n\n`searchDocuments` returns text passages with document names, page numbers, and relevance scores. Read the passages — they are actual document content, not just file names. Use them to understand what the data room says on the topic.\n\n### 3b — Keyword search for specifics (secondary)\n\nAfter the semantic search, run `searchDocuments` for any specific terms, figures, or exact phrases that the question calls for — e.g. a specific contract name, a company name, a regulation, a year, a metric. Keyword search complements semantic search for precise lookups.\n\n### 3c — Browse to the relevant folder if needed\n\nIf the search results point to a specific section of the data room but you need to confirm what documents are present (e.g. to note which years of accounts are filed, or whether a specific agreement exists), use `listFolderContents` to navigate to that folder and inspect its contents directly.\n\n### 3d — Synthesise and draft the answer\n\nWith the passages and document context in hand, write a clear, factual response. The standard to aim for:\n\n- **Directly answers** what was asked — not a broader essay on the topic\n- **Grounded in the documents** — reflects what the data room actually says, not general knowledge\n- **Sell-side voice** — professional, concise, confident. Written as if the CFO or GC reviewed it, not as a transcript of search results\n- **Handles uncertainty correctly** — if the data room contains partial information, say so clearly (e.g. \"Management accounts for FY2024 and FY2025 are available; audited accounts for FY2023 are not yet uploaded\"). Never fill gaps with assumptions.\n- **Sensitive matters** — if a question touches on active litigation strategy, unpublished projections, or personal employee data, flag it for legal review rather than drafting a response\n\n### 3e — Assign a status\n\n- **Complete** — question fully answered with clear source material\n- **Partial** — answer drafted but source material is incomplete or only partially responsive\n- **Open** — insufficient source material found; needs manual input from the deal team\n\n### 3f — Build the source reference and citation\n\nFor every answer, record two things:\n\n**Source Reference** (brief, for the tracker): the VDR folder path and document name — e.g. `3.1 Audited Accounts / FY2024 Annual Report` or `5.3 Customer Contracts / MSA with Acme Corp`\n\n**Document Citation** (detailed, for verification): the full citation including document name, VDR index path, and page number(s) where the relevant content was found — e.g. `FY2024 Annual Report (VDR 3.1), p.14 — Revenue recognition policy` or `Employment Agreement — J. Smith (VDR 7.2.4), p.3 — Clause 8, Non-compete`. If multiple documents were used, list each on a separate line.\n\nIf no source is found after running both semantic and keyword searches and browsing the relevant folder, mark the question Open and note: \"No source material found in data room — requires manual response.\"\n\n---\n\n## Step 4 — Group questions by theme\n\nBefore producing outputs, group questions into thematic sections. Common M&A Q&A groupings:\n- Financial Performance & Accounting\n- Tax\n- Legal & Regulatory\n- Commercial & Customers\n- Human Resources & Management\n- Intellectual Property & Technology\n- Operations\n- ESG & Environmental\n- Other / Miscellaneous\n\nUse the question content (and any category column already in the input file) to assign each question to a section.\n\n---\n\n## Step 5 — Offer outputs\n\nBefore generating the Excel tracker and dashboard, ask:\n\n> \"I've drafted answers for all [N] questions. What would you like me to produce?\n> - **Excel tracker** — formatted spreadsheet with all questions, answers, statuses, and source citations\n> - **Q&A management dashboard** — interactive React dashboard for active deal management (uses additional credits)\n> - **Both**\n> - **Neither** — just show me the answers in this conversation\"\n\nOnly generate the Excel tracker and/or dashboard if the user explicitly requests them.\n\n## Step 5b — Produce the Excel tracker (only if requested)\n\nUse the xlsx skill to produce a formatted `.xlsx` file saved to the outputs folder.\n\n**Columns (in order):**\n1. **Diligence Question** — the original question text verbatim\n2. **Diligence Response** — the AI-drafted answer\n3. **Status** — Complete / Partial / Open\n4. **Question From** — buyer group or individual name\n5. **Source Reference** — VDR folder path and document name (brief)\n6. **Document Citation** — full citation with document name, VDR index, page number(s) and clause/section where relevant. Multiple sources listed on separate lines within the cell.\n\n**Formatting rules:**\n- Header row: dark blue background (`#1a2332`), white font, bold\n- For each new theme/section, insert a **separator row** spanning all 6 columns containing the section name, styled with mid-blue background (`#2d4a6e`), white bold text — a visual divider, not a data row\n- Enable **text wrapping** on the \"Diligence Response\" column (column B) and \"Document Citation\" column (column F). Set column widths: B ~60 chars, F ~50 chars\n- Status cell colour coding: Complete = light green fill, Partial = light amber fill, Open = light red fill\n- Freeze the header row\n\nSave as `[ProjectName]_QA_Tracker_[Date].xlsx` in the outputs folder.\n\n---\n\n## Step 6 — Produce the React Q&A management dashboard (only if requested)\n\nRead `references/dashboard-spec.md` for the full React component specification before building.\n\nThe dashboard is a self-contained React component populated with the actual questions, answers, statuses, buyer groups, source references, and citations generated during the Q&A drafting process. It is for active deal management — it should feel live and usable, not like a static report.\n\nKey sections to implement (details in the reference file):\n1. **Summary KPI Bar** — four stat cards (Total, Open, Awaiting Review, Submitted)\n2. **Past Q&A Trackers** — collapsible card with drag-and-drop upload zone for precedent deals\n3. **AI Buyer Group Q&A Analysis** — collapsible panel with per-buyer stats, topic volume charts, and AI strategic signal\n4. **Filter Bar + Question Log** — searchable, filterable list with expandable rows showing AI draft, VDR citations, and management feedback thread\n5. **Dashboard Modal** — full KPI and analytics view with time-savings metrics\n\nUse navy `#1a2332` / gold `#d4a017` colour palette with Source Sans 3 font. All state via `useState` — no backend required.\n\n---\n\n## Step 7 — Deliver to the user\n\nPresent both outputs:\n1. Link to the Excel tracker file\n2. The React dashboard artifact rendered in the conversation\n\nThen say:\n> \"I've drafted answers to [N] questions — [X] Complete, [Y] Partial, [Z] Open. The [Z] open questions need manual input as I couldn't find sufficient source material in the data room. Both the Excel tracker and the live dashboard are ready above.\"\n\nIf there are Partial answers, offer:\n> \"For the [Y] partial answers, want me to flag the specific gaps so the team knows exactly what additional material to source?\"\n\n---\n\n## Operating principles\n\n**Read the documents, don't just locate them.** `searchDocuments` returns actual text passages — use them. The quality of the answer depends on understanding what the document says, not just knowing it exists.\n\n**Source everything.** Every drafted answer must have a citation. Unsourced answers should be marked Open. Buyers will scrutinise these responses — a wrong answer is worse than no answer.\n\n**Write in the seller's voice.** Concise, factual, professional. Not a summary of search results.\n\n**Don't over-answer.** Answer the specific question asked. Buyers will follow up for more.\n\n**Flag patterns.** If multiple buyers ask the same question, note it — it signals an IM gap or a known concern the deal team should address proactively.\n\n**Respect sensitivity.** Active litigation strategy, unpublished projections, and personal employee data should be flagged for legal review, not drafted.\n\n## Performance Notes\n\n- **Quality over speed.** A wrong answer is worse than no answer — buyers will scrutinise every response.\n- Read the source passages returned by `searchDocuments` fully before drafting. Do not skim.\n- Do not skip the keyword search step for questions involving specific figures, dates, or names.\n- Mark questions Open rather than guessing when source material is insufficient.\n\n---\n\n## Common Issues\n\n**`getProjectOverview` fails or returns the wrong project**\nCheck that the Datasite MCP connector is connected (Settings → Extensions → Datasite should show \"Connected\"). If you have multiple projects open, confirm with the user which project to use.\n\n**`listFolderContents` returns no results**\nThe fileroom may be empty or unpublished. Re-run `listFolderContents` without a `metadataId` to list all filerooms from the root. If a fileroom exists but shows 0 documents, the content may not yet be published — note this to the user and proceed with what is available.\n\n**`searchDocuments` returns an activation link instead of results**\nBlueflame AI search is not yet active on this project. Follow the Blueflame prompt in the skill instructions above. Do not attempt to answer using Claude's training knowledge.\n\n**MCP disconnects mid-workflow**\nReconnect via Settings → Extensions → Datasite. Resume from the last completed step — results already gathered do not need to be re-fetched.\n\n**`updateContent` or `createContent` returns a permissions error**\nThe user's Datasite account may not have Editor permissions on this project. Ask them to check their role in Datasite project settings.\n"
}