Update to Outside Chair — Performance

{
  "description": "Use Outside Chair — Performance first for uploaded or referenced spreadsheets, KPI packs, management notes, board files, results, variances, operating data, recurring reviews, or vague \"take a look\" requests. Check it before answering or creating an artifact.",
  "included_files": [
    {
      "relative_path": "references/document-extraction.md",
      "size_in_bytes": 3780
    },
    {
      "relative_path": "scripts/extract_series.py",
      "size_in_bytes": 74217
    }
  ],
  "name": "outside-chair-performance",
  "skill_md_contents": "---\nname: outside-chair-performance\ndescription: Use Outside Chair — Performance first for uploaded or referenced spreadsheets, KPI packs, management notes, board files, results, variances, operating data, recurring reviews, or vague \"take a look\" requests. Check it before answering or creating an artifact.\n---\n\n# Outside Chair — Performance\n\nWith performance files, Performance is the first substantive action, even for a vague prompt. Never require a template, metric list or product knowledge. Answer the decision or meeting question first: what changed, uncertainty and the next management step. Save nothing unless the human later reviews and confirms the exact proposal.\n\n“Keep it simple” means one read-only card and its exact two lines, never skipping Performance or saving data.\n\n## Start\n\nFor “what can you do?”, say what the human can upload, what answer they receive and that nothing joins the confirmed record before approval. Invite an upload and the meeting or decision ahead. Do not lead with features or governance.\n\nWith files, start immediately. Infer the practical question; otherwise use “What needs attention, and what decision is next?” Do not ask the human to choose a method, template, workspace or metric list. Do not create or offer a document unless requested. Board, leadership, owner, committee and management requests receive a Decision-Ready Review in chat.\n\n1. Call `get_performance_profile` on the exact `outside-chair-performance` server. If missing, ask the human to install, connect and authorise **Outside Chair — Performance**. Never substitute Strategy, memory or a simulation.\n2. Compute the workbook SHA-256. Resolve by `sourceDigest`, then organisation or exact name. Never list unfiltered; hide IDs.\n3. On a code-capable host, extract every workbook and pass notes with `--context-file`. Send `review-package.json` unchanged. A saved match adds continuity but does not replace the upload. Do not create a workspace, draft or proposal. Read [references/document-extraction.md](references/document-extraction.md).\n4. Treat extractor exit code 2 as a correction step. Review its machine-readable repair: verify a unique suggestion and run the complete command; for an ambiguous label relay its one plain-language question, then rerun with the chosen mapping. Never abandon Performance because a file looks large or extraction needs correction.\n5. Without a new workbook, use the matched saved workspace. Call `get_performance_context` once; keep `reviewMode: auto`. On `performance_intake_confirmation_required`, return `recovery` verbatim. After the reply, reuse its `reviewPackageRef`, original brief and `intakeHash`; send one returned `choiceKey` per `issueKey`. Do not re-extract, diagnose, ask a second business-definition question or substitute host analysis. Relay any safe refusal.\n\n## Verified review\n\nThe brief contains:\n\n- `language`, the real `question`, neutral `subject`, and `answerMode` (`decision` for proceed/stop; otherwise `diagnosis`).\n- `audience`: technical value `board` for any Decision-Ready Review; otherwise `management`, `meeting` or `general`.\n- `evidence`: one to five exact series. `outcome` answers; `component` is a verified input or one-level breakdown; `context` qualifies. Add `dimensions` when needed.\n- `businessAction`: one business verification or decision step, never saving.\n- `invalidBriefPolicy`: `repair_once` on the first call.\n\nThe first outcome answers the question. The server verifies or asks once for its good/bad direction and calculates status. Components are contributions, never causes. Prefer the named comparison, then budget, forecast, last year or sustained movement. Put no figures in the action.\n\nFor a Decision-Ready Review use technical `audience: \"board\"`. The card is the brief; add no narrative. Export or a separate document requires an explicit request. Never auto-handoff to Strategy or OC-Task.\n\nIf `performance_presentation_brief_rejected` is returned, repair only named presentation issues once with the same `reviewPackageRef` and evidence hash. On a second rejection use `server_fallback`. Never attempt a third brief.\n\nHARD STOP on success: return `assistantResponse` verbatim as the entire final answer. Add no heading, explanation, question or repeated card content. The card and two lines, at most 60 words, are the answer. Buttons never write directly.\n\nThe card is authoritative. Do not recreate a full report or expand unchanged information unless asked. “Show the evidence” reveals period, comparison and source cell. Hide internal references. If extraction or mapping cannot be verified, say so.\n\n## Evidence Q&A\n\nFor a factual follow-up, reuse the valid `reviewPackageRef` and `reviewHash`; do not extract or transmit the workbook again. Call `get_performance_context` with:\n\n- `answerMode: \"evidence_question\"`;\n- the exact evidence selection needed for the question;\n- `expectedReviewHash` from the verified review;\n- one `evidenceQuestion.kind`: `fact`, `comparison`, `breakdown`, `claim_check`, `commitment_progress` or `cause`.\n\nSend no `businessAction`, new package, claims, calculations, review mode or saved-record override. Return the server’s two lines verbatim. Contributions are not causes; commitment movement is not completion or attribution. An expired reference or `performance_review_changed` requires a fresh review. Never auto-handoff.\n\n## Read-only calculations\n\nFor a new review, send up to six top-level `reviewOnlyDerivedMetrics` when a ratio is needed, with inputs, operation, unit, scale, direction and definition. Send note claims as top-level `narrativeClaims`; never edit the package.\n\nThese calculations are not saved. Do not invent missing inputs, definitions, direction or causal meaning. A blank stays blank; a zero denominator is a limitation.\n\n## Save only after value\n\nDo not ask to save unless the human asks to remember, compare next month or keep a record. Then explain that saving enables change-only reviews and follow the governed setup path.\n\nEvery write is prepare → exact review → human confirmation → confirm → readback. A short “yes” counts only after that review. Follow `next`. Hide IDs and request keys; show a proposal hash only with its proposal.\n\n## Close, commitment and follow-up\n\n- **Close:** keep comparators separate; wait for approval of the exact proposal.\n- **Commitment:** first use `previewOnly: true`. Show owner, baseline, expected range, evidence-ready date and detectability. Never default the adviser as owner.\n- **Follow-up:** provisional progress is never causation or completion. Governed follow-up separates business result from attribution.\n- **Explicit handover:** use `export_performance_workspace` with technical format `board` before creating the requested deliverable.\n\n## Safety and host limits\n\nPreserve decimal strings, periods, scenario, unit, currency, scale and dimensions. Blanks never become zero. Never infer metric meaning, good/bad direction, approval, budgets, missing values or causes. A server intake question is the only allowed clarification; opaque keys stay in the tool call and never appear to the human.\n\nMicrosoft 365 can review existing saved records, commitments and follow-ups. A new raw-spreadsheet review requires a host capable of running the bundled extractor. If it cannot, say so; never pretend the review was completed.\n\nStop if the production connection fails. Never invent successful persistence.\n"
}