Writer Verification

Standing Mandates

• ALWAYS include original → fix + reason in every finding. Pointing out without fixing is half a review.
• ALWAYS run the humanizer pass. Correct grammar and good rhythm do not make text read as human; the tells are structural — headers on short text, what-without-why, nobody's-voice register.
• NEVER, in draft mode, hand over a draft that has not been through the passes. The first draft is input to the loop, never output.
• NEVER, in draft mode, add a header, bold label, or bullet the text did not need at that length. A PR description for a 40-line diff is three sentences.
• NEVER, in draft mode, hand over a prose wall either: why inside the first 30 words, no paragraph past ~80 words, and a flat list when there are three or more parallel concrete items. Reviewers skim; scannable is the human shape.
• NEVER write a PR description that lists files or opens with "This PR …". Open with why; the diff shows what.
• NEVER invent a fact while drafting or fixing. Every number, cause, name, and expanded acronym comes from the material or from the author — PG stays PG unless the material says what it stands for. A pass that "clarifies" a term it cannot verify has introduced an error, not fixed one.
• ALWAYS write in the language the user asked in, unless they say otherwise. A Korean request gets a Korean PR description, code identifiers untouched.
• ALWAYS stop the draft loop on 🔴🟡 = 0 or after three rounds, and say which — a fourth round is polishing, not fixing.
• Goal: a reader cannot tell this was reviewed or drafted by a model, and a reviewer of a PR knows where to look and what the author was unsure about.

Writer Verification

Five review passes over a text, each with one job, and a draft loop that runs them on Claude's own writing until the findings run out. The fifth pass is the one that closes the human-vs-AI gap: it does not check whether the text is correct, it checks whether a person would have written it that way.

Process

0. Decide the mode.

Also note the genre (pr · commit · doc · blog · message) and audience; the humanizer weights findings by genre.

1. Review mode. Count characters.

• < 300 chars — inline: read references/passes.md, run all five passes in order.
• ≥ 300 chars — parallel: dispatch all five in one turn, wait, aggregate.

Uncertain Korean spelling → references/korean-spelling.md. Genre pr/commit → references/pr-description.md is loaded by pass 5.

Aggregation: deduplicate by span, keep the higher severity and note both sources; conflicting fixes are shown side by side with attribution. If mcp-reasoner is available, use it when two passes flag one span for incompatible reasons.

2. Draft mode.

1. Gather the material. For a diff: commit messages, linked issue, the diff itself — derive why as references/pr-description.md describes; if why is not recoverable, ask the author one question before writing.
2. Write the first draft as the author would say it to a colleague — in the language the user asked in, at the length the material earns, no scaffolding.
3. Run the five passes on the draft (inline or parallel by length, as above).
4. Apply every 🔴 and 🟡. Rewrite the affected sentences; do not patch words in.
5. Repeat 3–4 until 🔴🟡 = 0 or three rounds have run. From round 2 on, re-run only the passes that returned 🔴🟡 in the previous round, plus the reader pass on any rewritten paragraph — a pass that came back clean stays clean. The exception is the material check: a rewrite that answers a [why] finding is where invented causes enter, so before the next round, read every "before/after" sentence the rewrite added against the diff (references/pr-description.md, "Reading a diff") and cut any claim without a - line behind it. Measured: the loop's own fix introduced "silently dropped" for a parameter the old code already forwarded. A finding only the author can close — missing doubt when the material states none, missing reason when it names none — moves to the "left for you" note after the round it first appears in and stops counting toward 🔴🟡; re-flagging it three times is what ran a 172-word draft to the round cap.
6. Deliver the final text, then a three-line note: rounds run, what the loop caught, what 🟢 remains for the author to decide.

3. Summary (review mode). If think-tool is available, use it here to pick the one or two findings that most change the reader's experience; those lead the summary.

Output Template

Review mode — full template in references/output-format.md:

## Spelling & Grammar · ## Writing Patterns · ## Expression & Style · ## Reader Perspective
## Humanizer
- 🔴 [why] "This PR refactors the retry logic" — what without why → "Checkout timed out for ~3% …"
- 🟡 [shape] 4 headers over 120 words → three sentences, no headers
## Summary
[2–3 lines: what works, the 1–2 changes that matter most]

Draft mode:

[final text — nothing above it]

---
Draft loop: N rounds · caught: [why missing, 4 headers → prose, 3× "leverage"] · left for you: [🟢 …]

Priority: 🔴 Must fix (meaning, logic, missing why in a PR) · 🟡 Recommended (patterns, tells) · 🟢 Optional (style preference).

What Claude Does / What You Do

Related Skills

• write:doc-coauthoring — when the document's content is still being decided, not its wording
• write:technical-blog-writer — long-form drafting; hand the result here for the loop
• develop:clean-code — for the commit itself; this skill covers the description of it