voice-transfer

name	voice-transfer
description	Rewrite a passage in a target author's voice using a corpus-calibrated critic. Use when asked to rewrite, retell or pastiche text in a specific author's style and a fitted handprint reference for that author exists.

Voice transfer

Rewrite a passage in an author's voice, using a fitted handprint reference as the objective function and the author's own corpus as the source of exemplars.

The reason this is a four-stage loop rather than a prompt is that the two known failure modes have opposite fixes. Mechanics converge easily and wit does not (GPT-4o with 15k-word excerpts matches sentence length and stays in its generic cluster on everything deeper). And imitation over-fires devices — the Dark & Stormy result — because a prompt gives you no target rate. So: plan the devices at the author's empirical density, render with mechanism-matched exemplars, and let the calibrated critic pull both ways.

Before you start

You need:

• REFERENCE — a fitted, calibrated reference for the target author. handprint pack show $REFERENCE should print a calibration line. Without one the pass gate cannot be evaluated and this skill has no stopping condition.
• SOURCE — the passage to rewrite.
• Optionally a codeindex corpus with a move index (see eval/agents/move-index/). Without it, skip stage 3's retrieval and say so in the report; the loop still works, less well.

If the reference is uncalibrated, stop and say so. Do not substitute "all findings cleared" for the gate — clearing every reported finding is exactly what an agent optimizing the feedback does, and the gate exists to catch it.

Stage 1 — plain retelling

Rewrite SOURCE with no style pressure at all. Preserve every event, every named entity, every stated fact. Aim for flat, unmarked prose.

This draft is the drift baseline: content_overlap is measured against it for the rest of the loop, so a stage-1 draft that already drops half the content makes the guard meaningless. Save it as plain.md.

Stage 2 — the device plan, at the author's rate

Read the target rates out of the reference:

handprint critique -r "$REFERENCE" plain.md --max-findings 200 > baseline.json

Every finding carries target_band. For each device dimension — dev:*, frame:*, syn:*, surp:punch_* — the band's midpoint is the author's empirical density. Convert it: a frame:negated_vehicle_rate band of 0.8–2.4 per 1k tokens over a 900-word passage means one or two negated similes in the whole piece, not one per paragraph.

Write the plan as a list of beats. For each:

1. Which beat of the source it covers.
2. Which device, if any. Most beats get none — that is what the rate says.
3. A remote association step for the ones that do: what distant domain can this be yoked to? Bureaucracy for a cosmic event, domestic detail for a catastrophe. Generate three candidates and pick the least obvious.
4. The punchline written out as a one-liner, before any prose exists.

A plan that puts a device in every beat has already failed. Check it against the rates before writing a word.

Stage 3 — render

Per scene, retrieve 2–4 exemplars by mechanism:

codeindex search "$AUTHOR: character introduction using register clash" --limit 4

Not by topic. Topical exemplars convey register and not technique, which is the finding the whole move index exists to act on.

Imitate the mechanism, not the sentence. The exemplars are there to show how the device is built; a phrase lifted from one is parroting and the drift fingerprint will catch it. If you notice yourself reusing an exemplar's vehicle, that is the signal to invent a different one.

Stage 4 — the polish loop

handprint critique -r "$REFERENCE" draft.md > report.json; echo $?

Exit 0 pass, 1 findings, 2 error. Iterate up to 12 times, and sequence the work:

• Iterations 1–4: mechanics. Act on punct:*, sent:*, mfw:*, read:* findings. These converge fast and there is no point judging wit through the wrong sentence rhythm.
• Iterations 5+: wit. Act on dev:*, frame:*, surp:punch_*, form:*. Run the wit critic (eval/judge/) if one is available.

Read the whole report each time, not just findings:

• guards.canary_ok == false — stop patching. You have been moving the reported numbers while the held-out family got worse, which is the definition of optimizing the feedback. Rewrite the passage as prose from the plan instead of editing the flagged items.
• guards.drift_ok == false — the rewrite has left the content behind. Go back to plain.md and re-render the beat that drifted.
• doc.confidence — a family reading none at this length is noise. Do not act on its findings.
• direction: "increase" is as real as "reduce". Under-firing a device is a mechanics-only pastiche that never acquired the voice, and the ○ cells of the coverage matrix are precisely this claim. An author who never uses adverbs has a low band on adverbs, and going below it is as much a miss as going above.

The gate

A draft is done when all of:

1. verdict.pass == true — p_same_author at or above threshold. Not "all findings cleared".
2. guards.canary_ok && guards.drift_ok.
3. Zero Medium-or-above findings in either direction on Device, Frames, Syntax and Rhythm dimensions. Inside the band, not merely under the ceiling.
4. If a wit critic is available: its pairwise win rate against gold anchors improved over the stage-1 baseline.

Each of those alone is Goodhart-able. The ensemble is the gate — that is why there are four and not one.

Report

State plainly:

• Which arm ran (plain-only, critique-loop, full pipeline).
• Iterations used, and whether the gate was met or the cap was hit.
• Final p_same_author, content_overlap, and any guard that tripped.
• Which device dimensions ended outside their band, in which direction.

If the loop did not converge, say so and show the last report. A voice transfer that hit the iteration cap is a result, and writing it up as a success is the one failure this whole design exists to prevent.