| name | style-mimicry |
| description | Reproduce an author's writing voice when drafting reports/updates from a few bullet points, using their past work as reference. |
| when_to_use | the user wants something written in their own established style (daily/weekly reports, updates) rather than a generic AI tone |
Style Mimicry
Goal: make the draft read like the user wrote it, not like a generic AI.
Workflow
- Confirm the clue. Treat the user's message as the factual seed for this
piece. Confirm it is the clue before drafting; if it is thin, ask for the
key facts you must include — never invent details.
- Pull references. Retrieve the user's most similar past pieces (via the
agent's retrieval tool / long-term memory). These define the voice.
- Extract the voice signature from the references before writing:
- typical length and paragraph rhythm
- sentence length and formality
- recurring connective phrases / openings / closings
- how numbers, jargon, and hedging are handled
- Draft matching that signature. Hard rules:
- Keep length close to the references — do not balloon a 5-line update into an essay.
- Use only facts from the clue or references. No new claims, no invented metrics.
- Avoid telltale AI tics (over-structured headings, "In conclusion", excessive enthusiasm) unless the references use them.
- Hand off for editing, then offer to persist the approved final version
back to memory/the database so future drafts get closer to the user's voice.
Anti-patterns to avoid
- Generic corporate filler the user never uses.
- Inflating scope or certainty beyond the clue.
- Restructuring the user's natural format into a template.
