Rewrite

Two ways in, and they differ only in what tells you where the text is wrong:

Why it is rounds and not a pass, and the measurements behind every rule here: loop.md and measurements.md. Neither is needed to act; this file is.

Step 1. The measurement, where there is one

Ask the user for the run directory from audit and read audit.md there. You need two headings: Reader profile and What broke. The contract is in ledgers.md.

Step 2. The brief

Assemble it once, for every writer, in this order; a resumed run does not reassemble it. On the skeleton route parts 2 and 5 do not exist; say so in the report rather than inventing them.

1. The skeleton, if there is one — its purpose, exclusions and budget for the section being written, unedited.
2. Who reads this — the profile from the run file, unedited.
3. The writing rules — writing-rules.md, copied in as written.
4. The curse of knowledge — curse-of-knowledge.md, likewise.
5. Where the readers failed — the entries under What broke, each with its line, its cause and its quote, plus the passages that worked and must not be damaged.
6. The accuracy floor: every statement about behaviour must be true of the code in this checkout, and the writer records the level of evidence it reached — the three levels are in truth-pass.md. A claim about a lifecycle (what stays, what is removed, what a continued or retried run sees) at level 2 is a guess: run it.

Copy the fixed parts; do not paraphrase them. The wording a critic receives is in bake-off.md; do not write a third.

Step 3. The first candidate

On a document that already exists, the first thing that runs is the adversarial whole-document read — lens 3 of step 4's table — with the right to run the code. On the audit route its findings are added under What broke; on the skeleton route they become the first round's edits. Then the bake-off: three writers, one whole candidate each with a different stance, and two judges — the briefs and the judging sheet, for both routes, are in bake-off.md. That is the only bake-off; every round after it edits the round before.

Announce before spawning: the count, the models, and that the cost of a writer or a judge has not been measured (the critics' costs have; see the table). Wait for the user's word. If the user refuses the fan-out, write one candidate yourself from the same brief and report that the comparison was skipped.

Step 4. The rounds

Work in a run directory of the document's own — research/<date>-<slug>/ at the root of the repository that holds the document; audit writes its run file elsewhere, and you copy audit.md in. Every round is its own file, NN-<pass>.md, named for what produced it — 01-candidate, 06-water, 08-review — any name, used once; the original is 00-original.md. Set S to this skill's scripts/ directory once, as an absolute path: installed, it is $CLAUDE_PLUGIN_ROOT/skills/rewrite/scripts; from a checkout, the scripts/ beside this file. Then:

node "$S/selftest.mjs"      # once per session: every check against its planted violation

Every command below is written for the run directory as the working directory; pass absolute paths if your shell rules forbid cd.

1. Once per document, write four files and copy in skeleton.md:
   • concepts.json — one regex per idea the document carries, for the duplication count;
   • budgets.json — every ## heading of the document mapped to the skeleton's number, or one you set and write back into skeleton.md where a section was added later. The comparison is a report the owner reads, not a gate: growth per section per round is the number that showed four sections swelling while every fix made them truer;
   • tasks.json — two starting states and goals for lens 4, from the workflow the document most wants a reader to perform; questions.json — the questions a reader arrives with, from the audit's key when there is one, otherwise from the skeleton's purpose per section, one line each. On a resumed run reuse all four; the readers are new agents and stay fresh even where the questions repeat. ledger.json starts empty and grows from the rounds.
2. Write edits/NN.json: for each edit the exact old text, which must occur once; the new text; claims it introduces and retire phrasings it removes as false; and check — {"level": 1|2|3, "how": "command or file:line"} — required whenever the edit carries claims. The format is the header of round.mjs.
3. Produce the round: node "$S/round.mjs" <NN-1>-<pass>.md <NN>-<pass>.md edits/NN.json --ledger ledger.json.
4. Run the checks, before any critic, with R=<NN>-<pass>.md:
   • node "$S/rule1.mjs" "$R" --cut "<technical section heading>" --except "<section that may carry paths>" — the rule that keeps mechanism out of the sections a reader meets first, with the document's own headings;
   • node "$S/dup.mjs" "$R" concepts.json — one idea, one home;
   • node "$S/sections.mjs" "$R" budgets.json — words per section against the budget, reported, never blocking;
   • node "$S/ledger.mjs" ledger.json $(ls [0-9][0-9]-*.md | sort) — the ratchet over every round in order; exit 1 when the new round loses a verified claim or revives a retired phrase. A failure the round introduced is fixed before the critics see it: remove the round file, fix edits/NN.json, regenerate. A failure the previous round already had is a finding for this round's edits, not a block. A round is frozen the moment its critics launch, not before.
5. Announce the wave — the lenses, their sizes from the table, the models, the cost — and wait for the user's word; the user may size any lens to zero, and the least that still counts as a round is lenses 1 and 2. Then launch the critics, one agent per lens, with the briefs in critic-briefs.md, and the dedup agent over their reports; everything they return is kept verbatim under reviews/NN/. The Codex lenses need the entrust plugin; without it, run those lenses on Claude agents and say so.
6. Verify every finding yourself from the check it carries — a finding without one is discarded — and route each by the table in loop.md: a sentence to the next round's edits, a boundary or a term back to rethink, a code defect to the repository's ISSUES.md, a question the document does not answer to the user. When a finding routes to stage 3, make the structure map loop.md describes before the next round.
7. Record the round in rounds.md, one table row: | file | words | produced by | findings against it | regressions | — produced by names the pass and the wave; findings is the dedup's count by category; regressions is the count of sentences the round introduced that its critics showed false or overstated. That number is the round's verdict.

The lenses are fixed; the sizes are the user's, and the announcement names them. Costs are what one wave measured on 2026-09-12 on a 1600-word README:

Lenses differ; they are not disjoint, and a finding three of them raise is confirmed, not counted three times.

Step 5. The gate, and the stop

Before the user reads a round, three things, none tradeable against another:

• no regression in the round: the ledger passes, and no sentence the round introduced was shown false or overstated by its critics;
• the task gate: lens 4's two readers achieved their goals, and the sections no task reached are named;
• the question readers: lens 5, one per question, answered from the document; where they guessed is listed.

The loop stops when the user reads the round and says whether they would send it as it is. Two consecutive rounds with no regression is the signal to hand a round over, not a finish; a cap on rounds is set in the first announcement, and a cap reached is reported as a result. Hand over the round and diff-NN.patch, the diff against 00-original.md, written into the run directory. Then stop: applying the candidate to the user's files needs their word, and a diff they have read is what earns it.

Step 6. What you return

In the run directory:

• every round as its own file; edits/NN.json, ledger.json, concepts.json, budgets.json
• skeleton.md, kept current with every decision taken after it was agreed — a section added, a fact restored, a budget changed
• rounds.md, one row per round; reviews/NN/, every critic's report and the dedup, verbatim
• diff-NN.patch against the original; the cut ledger — every removed passage of twenty words or more, with its reason; the invisible-prerequisite inventory from the curse-of-knowledge pass, on rounds that ran a writer brief; the sections no task reached; and the structure map, when a stage-3 finding called for one

Formats for the ledgers are in ledgers.md.

The safeguards are the last lines of writing-rules.md and override the rest of the rules wherever they collide: never cut a condition, a limit or a warning where a reader decides; repetition at an independently read decision point is not redundancy; a dated measurement keeps its date and its numbers.

Reference

• Why rounds, the routing of findings, the ledger, the map and the checks: loop.md.
• The measurements behind every rule here, dated: measurements.md.
• The three decisions before any sentence: stages.md.
• Writer briefs and judging sheets, both routes: bake-off.md.
• Critic briefs, one per lens, and the rights header: critic-briefs.md.
• The checks as scripts with a planted-violation self-test: scripts/.
• The rules, fixed: writing-rules.md; the third pass, fixed: curse-of-knowledge.md.
• Run file and ledger formats: ledgers.md; evidence levels: truth-pass.md.