Dossier Intelligence

Path B: Incremental Update

Run this when the user wants to sync the index with filesystem changes.

1. Execute:

   python ${SKILL_DIR}/scripts/update_index.py "<corpus_path>"
   

2. Report: files added, updated, and removed
3. If all counts are zero, tell the user the index is already up to date

Path C: Search / Query (Default Path)

Follow this sequence strictly for every document query. Never skip to file reads.

Step 1: Search the index

python ${SKILL_DIR}/scripts/search_index.py "<user_query>" --corpus "<corpus_path>" --limit 10

If the user mentions a specific category, add --category <name>.

Step 2: Evaluate results

Review the ranked dossier list and their 3-sentence summaries.

Step 3: Branch on result type

• Summary sufficient: Answer the user's question directly from the index results. Zero file reads. This is the preferred outcome.

• Need 1-3 specific dossiers: Open ONLY those files with the Read tool for detailed analysis. Never open more than 3 files in a single query.

• Cross-corpus reasoning required (comparing 5+ dossiers across categories): Spawn one subagent per relevant category. Each subagent receives only its category's search results. Main session merges the structured summaries from each subagent.

Step 4: Respond

Present findings to the user. Always cite the dossier name and category. If answering from summaries, note that a deeper analysis is available if needed.

Rules

1. ALWAYS search the index before opening any dossier file
2. PREFER answering from summaries over full-file reads
3. NEVER open more than 3 dossier files without subagent delegation
4. If the index is missing, tell the user and offer to run Path A
5. Never read script source code into context
6. Report status messages in French when the user communicates in French
7. If the query returns zero results, suggest broadening the search terms or checking if the index needs updating (Path B)

Edge Cases

• If the corpus path does not exist: ask the user to verify the path
• If a file fails to open: report the error and continue with remaining files
• If the user asks about a file type not in the index (e.g., .xlsx): explain that only PDF, DOCX, and images are currently indexed
• If search returns too many results: ask the user to narrow with a category filter or more specific keywords

Reference

• Index format details: resources/index-schema.md
• Query routing logic: resources/query-router.md
• User-facing messages: resources/ux-messages.md