Restyle

restyle

Rewrites a Markdown draft toward a target author's voice, then quality-checks and repairs the result.

Usage: /restyle <draft path> --voice <name> [--contrast]

Paths

$SKILL below means the base directory for this skill, given to you at invocation. Every script, prompt and voice lives inside it. The skill is normally installed at ~/.claude/skills/restyle, so it runs from any project and the working directory is wherever the user's draft is — never assume the two are the same. Always invoke scripts by their full $SKILL/... path.

The draft path and the output files are the exception: those are relative to the user's working directory, because they are the user's documents, not the skill's.

If no voice is given, list the directories under $SKILL/voices/ and ask which to use. If the draft path is missing, ask for it. Do not guess either.

Inputs

• "$SKILL/voices/<name>/exemplars.md" — sent to the generator and the voice reviewer.
• "$SKILL/voices/<name>/corpus/" — never sent to a model; used only by the overlap check. .txt files are read recursively, so a corpus may be organised into subdirectories.

{{EXEMPLARS}} is substituted with the path, not the file's contents: the subagent reads it. Do not inline the exemplars — that spends thousands of tokens of your own context on text only the subagent needs. {{DRAFT}} and {{REWRITE}} are substituted with contents, because you are already holding them.

If exemplars.md is missing, stop and say so. If corpus/ is missing or empty, warn that the plagiarism check will not run, and ask whether to continue.

Preflight — always run this first

python3 "$SKILL/scripts/build_exemplars.py" --check "$SKILL/voices/<name>/exemplars.md" --corpus "$SKILL/voices/<name>/corpus"

Exit 1 means passages in exemplars.md are absent from corpus/. The generator would then be shown text the overlap check cannot detect it reproducing — the guard has a hole exactly where borrowing is most likely. Report which passages, and recommend voice build before continuing. Do not proceed silently.

Procedure

1. Generate

Read $SKILL/prompts/generate.md. Substitute:

• {{DRAFT}} — the draft file's contents
• {{EXEMPLARS}} — "$SKILL/voices/<name>/exemplars.md"
• {{CONTRAST_BLOCK}} — empty string by default. With --contrast, insert: Also state how the target voice differs from the draft's own style on each dimension, and let that difference drive the descriptors.

Dispatch one Agent tool call (general-purpose) with the substituted prompt. Generate exactly one candidate — review and repair will fix what is wrong with it, so more candidates buy nothing.

Extract the ## Descriptors and ## Rewrite sections. Write the rewrite to a working file. If either heading is missing, re-dispatch once; if it fails again, stop and report.

Default to target-only register analysis. --contrast exists for targets close to the author's own register; a distinctive literary voice is not one.

2. Overlap check

python3 "$SKILL/scripts/overlap_check.py" <working file> --corpus "$SKILL/voices/<name>/corpus"

Exit code 1 means matches were found. This does not halt the run. Convert each match into a blocker finding in the schema from $SKILL/prompts/finding-schema.md, using the matched text as the quoted anchor.

Exit code 2 means the corpus could not be read — missing directory, no .txt files, or nothing long enough to check against. That is not a clean pass and must never be treated as one. Stop and report it: the guard did not run.

The check uses an 8-word shingle and will not flag shorter echoes. That is a deliberate limit — a three-word rhetorical move is imitation, not reproduction, and catching it would mean flagging ordinary English. Do not lower -n.

3. Review

Read $SKILL/prompts/finding-schema.md once and substitute it into {{SCHEMA}} in all three reviewer prompts.

Dispatch three Agent tool calls in a single message so they run in parallel:

The withholding is deliberate and must be preserved. A fidelity reviewer who knows the target voice starts excusing invented content as style. A voice reviewer who has seen the draft starts rewarding timidity.

Collect all findings. Discard any finding whose quoted anchor does not appear verbatim in the rewrite — an unanchored finding cannot be verified or repaired. Record how many were discarded for the report.

4. Repair

If there are no findings, skip to step 5.

Read $SKILL/prompts/repair.md, substitute {{REWRITE}}, {{FINDINGS}} (all findings from all sources, most severe first), {{DRAFT}}, {{EXEMPLARS}}. Dispatch one Agent tool call.

Extract ## Repaired as the new working text and ## Unresolved for the report.

5. Iterate

Repeat steps 2-4 at most twice total. After the second repair, ship whatever you have.

Stop early if a round produces no findings.

Do not run a third round. Repair passes erode untouched prose, and a finding that survives two targeted attempts needs the author, not another attempt.

Round 2 matters even when round 1 looks clean: repair can introduce drift of its own, and re-review is what catches it.

6. Write output

Write the final text to <draft>.restyled.md.

Write <draft>.report.md:

# restyle report — <draft name>

- **Voice:** <name>
- **Variant:** RG (target-only) | RG-Contrastive
- **Date:** <YYYY-MM-DD>
- **Words:** <draft count> → <output count>

## Descriptors

<the generator's descriptor list>

## Round 1

- Findings: N blockers, N major, N minor (N discarded for missing anchors)
- Overlap matches: N
- Unresolved after repair: <titles, or None>

## Round 2

<same shape, or "Not run — round 1 came back clean.">

## Unresolved findings

<full text of anything still outstanding, or None>

## Overlap matches outstanding

<any still present after the final round, or None>

Then tell the user: output path, report path, and — stated plainly, not buried — any outstanding overlap matches. That is the one class of issue they should never publish without having looked at themselves.

Building a voice

/restyle voice build <name> regenerates "$SKILL/voices/<name>/exemplars.md" from the corpus. Run it when a voice is new, when corpus files change, or when preflight fails.

1. Extract candidates

python3 "$SKILL/scripts/extract_candidates.py" --corpus "$SKILL/voices/<name>/corpus" -k 12 --out /tmp/<name>_candidates.json

Returns the medoid passage of each register cluster, allocated per source file. The per-source allocation is load-bearing: clustering alone lets the largest file take most slots, because it holds most of the passages. Do not replace it with a flat sample.

2. Curate

Read $SKILL/prompts/curate-voice.md, substitute {{CANDIDATES}} with the candidates file path, and dispatch one Agent tool call. It returns JSON with selections (8-10 ids and register labels, ordered calmest first), rejected, and notes.

Write that JSON to /tmp/<name>_selections.json verbatim. Step 3 reads it from there; it is not passed to the script any other way.

Read notes. If it reports a register missing from the corpus — most often the calm expository end — tell the user plainly. That is a gap in their source material, not something to paper over, and it is what causes pastiche.

3. Assemble and verify

python3 "$SKILL/scripts/build_exemplars.py" \
  --candidates /tmp/<name>_candidates.json \
  --selections /tmp/<name>_selections.json \
  --corpus "$SKILL/voices/<name>/corpus" \
  --voice <name> \
  --out "$SKILL/voices/<name>/exemplars.md"

This writes the file and re-checks the invariant. A non-zero exit means the build failed; do not use the result.

Notes

• The descriptor list in the report explains the whole direction of the rewrite. When output is wrong, read it first.
• Never chunk the article for generation. It goes to the generator whole.