voice-matchinglisted

# Voice Matching ## When to Use Use this skill for any prose output longer than approximately 150 words that will be presented as the researcher's own writing. Canonical triggers: - Grant sections (Specific Aims, Significance, Innovation, Approach, Broader Impacts, Data Management Plan narratives) - Cover letters for manuscript submissions or job-search packets - Response-to-reviewer letters - NIH progress reports and administrative supplements - Lab-website text (research descriptions, people pages, news items) - Manuscript Discussion or Introduction sections drafted for review **Do not use for:** - Chat replies and tool-output summaries (these should be efficient, not voice-matched — they go to the researcher, not an external reader) - Internal notes, database comments, or annotation fields - Bullet-point action items or todo lists - Short confirmations ("done", "here is the list you asked for") The test: will this text be read by someone external to the lab as the researcher's own words? If yes, use this skill. --- ## The Voice Index `agent/voice_index.py` maintains a TF-IDF retrieval index of the researcher's published prose. The primary source is `data/voice_passages.json`, which contains curated paragraph-level passages from Discussion sections, grant narrative paragraphs, and lab-website prose. When the curated file has fewer than 40 passages, the index is supplemented from OpenAlex abstracts for the researcher's publications (fetched via the OpenAlex API u