AI agent skill · by @amd
Document Extract
Enumerate every requested item in long text documents or transcripts, with fields and source evidence. Use for exhaustive inventories such as every exercise, action item, or finding; not ordinary summaries, general knowledge lists, or code symbols.
Archived 2026-10-23 · Source updated 2026-10-12 · License: MIT. Check the original repository for the latest version.
SKILL.md instructions
# Exhaustive document extraction 1. Identify each source text file and the requested item type and fields. Use the user's current files; remembered summaries can help locate them but do not establish current contents or completeness. 2. Call `extract_document_items` once per source. It processes bounded chunks sequentially and retains source-backed entries. Do not replace this with a summary or manually calculated page offsets. When specifying fields in the tool call, pass `fields: ["name", "owner", "deadline"]` (the actual requested names). 3. If an extraction fails, report it as incomplete. Do not silently switch to a shorter list, a guessed count, or a script that has not been run. 4. If the user requested a saved inventory, use `save_extracted_items` for the exact destination. The exporter preserves all entries and provenance as JSON, CSV, or readable text, and the framework reads the file back and checks it; don't edit or regenerate it by hand. 5. The framework appends the full retained inventory to your final answer. Briefly state limitations or requested analysis; do not repeat or re-summarize the list. The existing memory system records the action, source fingerprint, coverage, count and final inventory unless the session is private. Use conversation search for follow-ups about previous work; re-extract if the source or requested fields change. Never treat a remembered summary as proof that all current items were extracted, and don't store a whole private transcript as a durable personal fact. A compact summary may supplement the inventory, not replace it.