START WITH THE PRODUCT STEPS Read START-HERE.txt first. It states the actual source acquisition, access/privacy limits, manual field mapping and exact next prompt. The original file templates and synthetic examples below are supporting review aids. Some paths need separately supplied documents or source text; a blank worksheet is not a native export or completed integration. S01: Find approved knowledge with source references Starter kit Status: Proposed / not tested PURPOSE An internal research/reference question arrives → search the approved knowledge index by topic and document type → a shortlist of passages, source IDs, versions and unresolved gaps. PREREQUISITES 1. Existing approved DMS or SharePoint library; a curated index of public or deidentified reference materials; approved taxonomy; current/superseded status and owner. Start with manual search and exported metadata. AI only summarizes retrieved approved material. 2. Synthetic files first. Real use requires firm-approved access, privacy, processing terms, retention controls, export coverage and a named reviewer. 3. Prepare any required original files or rendered evidence separately. This kit does not configure accounts, connectors or live actions. FILES AND FORMATS knowledge-index.csv: document_id, title, topic, version, approval_date, current_status, owner, source_locator approved-excerpts.txt: document_id, page_or_section, exact_text question.txt: bounded reference question Files without .example are blank input worksheets; CSV templates contain headers only. .example files contain invented records for the dry run. Text input .txt files contain blank labeled fields. Use the original base filename when supplying an example to the prompt, or explicitly map the example filename to that input. CSV files are UTF-8 with comma delimiters, quoted fields where needed, and CRLF row endings. Preserve identifiers as text when importing into Excel. Blank means missing or unresolved, never a zero or a pass. Multi-value fields use a pipe only where the fixture scope says so. SETUP AND DRY RUN 1. Copy this kit to an approved working folder. Keep examples separate from any real exports. Do not upload client, privileged, personal, credential or financial-account data to an unapproved service. 2. Read the .example text policy/scope files and all .example.csv inputs together. They define this fixture only. For a real pilot, fill the blank templates from approved exports without changing headers and obtain approved policy values. 3. Use native read-only exports, exact joins and arithmetic for deterministic checks. This kit includes no executable automation, OAuth setup, live-action instructions or importable Make blueprint. 4. Give only the approved prepared inputs to a firm-approved text assistant and use the recipe prompt below. Keep source IDs, source row numbers and unresolved evidence visible. 5. Compare the proposed result with expected-fixture-result.txt. Verify every listed issue and that the prompt invents no missing evidence. The expected file describes assertions, not a performed workflow run. 6. Have the named reviewer check source evidence and record remaining questions before any separately authorized action. RECIPE PROMPT Search only this approved reference set for the question below. Return the most relevant passages with exact document IDs, page/section, version and approval date. Explain why each might help. Distinguish an internal note from an authority. If no source supports a claim, write ‘not found.’ Do not give legal conclusions, update a precedent, or invent citations. Flag superseded and conflicting sources for attorney review. INPUTS knowledge-index.csv: document_id, title, topic, version, approval_date, current_status, owner, source_locator approved-excerpts.txt: document_id, page_or_section, exact_text; question.txt: bounded reference question OUTPUT knowledge-review.csv: question_id, document_id, version, approval_date, locator, exact_passage, relevance_reason, status, gap Use only these synthetic or firm-approved inputs and preserve source row/page IDs. Keep missing or conflicting evidence unresolved. Return a draft for the named reviewer. Do not send messages, connect accounts, change records or permissions, move money, or make legal decisions. Keep source files unchanged. List missing prerequisites before continuing. REVIEW CHECKLIST [ ] Headers match the exact input schema and the export covers the stated scope [ ] Synthetic example files were used alone for the first dry run [ ] Every output finding retains its original source ID/row/page reference [ ] Missing keys, contradictions, unavailable originals and ambiguous dates remain unresolved [ ] Every expected fixture finding is present; no live action or unsupported result is claimed [ ] Responsible reviewer checks the original records before a real deployment RECIPE-SPECIFIC REVIEW Knowledge owner validates tags and access; attorney verifies suitability and current law separately. Plant an outdated note and a question with no answer; neither may become an authoritative answer. IMPORTANT FIXTURE LIMIT The excerpts are internal synthetic reference notes, not legal authorities. A question ID is written into question.txt because the output asks for question_id. VALIDATION Exact CSV header order, field counts, serialized round-trip values and fixture semantics were checked during kit preparation. This does not test OCR, a vendor integration, a model response or end-to-end reliability. All data and identifiers are synthetic. Dates are fixed fixture dates, not current monitoring results.