name: voice-extractor
description: Read a corpus of the user's published papers (PDFs) and produce a VOICE.md style guide capturing recurring phrasing, sentence rhythm, hedging style, structural moves, and a ban list. Triggers when the user asks to "extract my voice", "build a voice file from these papers", or runs /voice-extractor.
args:
papers_dir: directory of PDFs (default: sample-papers/)
output: output path (default: output/VOICE.md)
voice-extractor
A four-phase skill. Each phase ends with a checkpoint the user can interrupt.
Phase 1 -- discover
List PDFs in papers_dir. If fewer than 3 files, stop and ask for more -- voice extraction needs a corpus, not a single sample.
For each PDF, read the prose body. Skip the references section, appendices, and the contents of tables -- voice lives in the running text. Note which sections each paper has (intro / data / methodology / results / robustness / conclusion) so Phase 2 can sample from comparable slots across papers.
Phase 2 -- extract patterns
For each paper, jot (in working memory, not on disk):
- 3 representative opening sentences of major sections
- 3 representative transitions or hedges (e.g., "Importantly, ...", "We caution that ...")
- 1 example each of: introducing a new result, qualifying a claim, citing prior work, framing the contribution
- sentence-length distribution (qualitative: short / medium / long, with a rough ratio)
- notable diction: favored verbs, recurring phrases, characteristic qualifiers
If a paper is co-authored with very different voices, note that and weigh it less.
Phase 3 -- synthesize
Write a single VOICE.md with these sections:
- Stance -- first/third person, hedging level, formality, audience
- Sentence rhythm -- typical lengths, parallel-construction habits, paragraph openers
- Recurring phrases -- annotated examples, each tagged with the paper of origin
- Structural moves -- how openings / contributions / robustness paragraphs are framed
- Diction preferences -- favored verbs, qualifiers, transitions
- Ban list -- words/phrases that don't appear (negative space matters)
- Worked example -- one short paragraph rewritten in this voice
Phase 4 -- verify
Re-read the VOICE.md you just wrote. For each entry under Recurring phrases, confirm that phrase actually appears in at least 2 of the source PDFs (use Grep over the extracted text). If a claim fails, drop or weaken it.
In the final message to the user, list any phrases that were dropped or weakened in this verify step. This is the pedagogical point of the skill -- the gate must be visible.
