Writing Review

One job: tell Shane what is wrong with a piece, biggest problem first, without changing a word of it.

The rule that makes this skill work

Never produce a replacement draft. The moment you write a competing version, the review stops being a diagnosis and becomes a negotiation about your prose. Shane invokes this when he wants to decide, not to be handed a rewrite. If he wants the edit, he will run humanize.

Also: never edit the file, never save a new file, never stage anything. The review is reported in chat and nowhere else.

The verdict vocabulary

Use exactly these four, and no others:

ask-author is what earns this skill its keep. It is the alternative to inventing the missing detail, which is the single easiest way to damage a draft while believing you improved it. A review with four ask-author findings is worth more than one with four clever rewrites.

Order of work

Diagnose in this order. Do not jump to sentences before the earlier layers are settled: a sentence-level finding on a piece with a substance problem is noise.

1. Truth and scope — claims, numbers, attribution, uncertainty, conditions, links.
2. Substance — the actual point, and the material only Shane could supply.
3. Development — does each paragraph depend on the one before it?
4. Sentences — surface patterning.
5. Craft — did the piece lose warmth, stance, or a landing?

Piece-level diagnosis

Open with this. Use the fields that carry information and omit the rest.

Job:          what kind of piece this is, who it's for, what should change for them
Substance:    what it specifically contributes, or what's missing
Trust:        factual support, attribution, uncertainty, and any preservation risk
Development:  how it moves, and where it stalls, repeats, or turns into a list
Voice:        what reads as Shane's, and where generic register takes over
Ending:       whether it stops on the last useful thought or reaches for a close
Top fixes:    the two or three changes with the largest expected gain, ranked

Passage-level findings

One block per material finding, ordered by Top fixes rather than by position in the draft.

Passage:      the shortest quote that locates the issue
Verdict:      keep / revise / ask-author / cut
Pattern:      a short name
Why:          what the passage does instead of its job
Suggestion:   a supported replacement, a precise question for Shane, or the reason to cut
Safety check: does the suggestion preserve facts, attribution, scope, and voice?

Before proposing any replacement, check it for the same formula, inflated claim, or invented detail you just flagged.

Tests to run before naming a finding

Substance test. Could this paragraph appear nearly word for word in someone else's article on the same subject? If yes, it is filler even if it is well made.

Flatten test. State the claim without the cadence. If nothing specific remains, the formula was carrying reasoning that was never done.

Relation test. Restate the implied connection using because, although, once, if, so, or which. If doing so requires inventing a relation, only rhythm was there.

Specificity test. Remove the specific. Does the argument weaken? If not, it was a flex, not evidence.

Triad test. Does each of the three items carry distinct weight? A parallel-grammatical three with identical structure is filler dressed as rhythm.

Do not report these as defects

Profiles/voice.md is authoritative for voice. These are deliberate choices, and a review that flags them is wrong:

• A fragment sequence that lands a build-up (Three emails sent. One reply. Zero closes.).
• The closing punch that restates the thesis in compressed form.
• The confessional or first-person opener.
• Hedges doing honest work, and the absence of a hedge where the claim is settled.
• Functional specificity: proper nouns, versions, file paths, gateway names doing real work.
• Headers carrying the argument, so a header-only read still surfaces the spine.
• One grounded colloquial aside, or none at all.
• A deliberately short sentence landing after a long build.

The governing calibration: change repeated use or unsupported performance, not the device itself.

Hard refusals bind the review

Profiles/voice.md's <hard_refusals> are absolute, so a violation is a finding no matter how well it reads:

• em-dashes anywhere, in any form
• "simply", "just", "obviously", "actually"
• em-dash lists, single-noun colon-reveals, semicolons between independent clauses
• highlight-reel structure, emoji-prefixed bullet headers
• top-ten lists, meme formats, regurgitated AI-session pastes
• invented specifics, war stories, or receipts

Reporting a clean piece

If the prose already performs its job, say so plainly and stop. Do not manufacture findings to look useful. "This is ready" is a valid review outcome and often the correct one.

Handoff

Close by collecting every open ask-author item into one list of questions. That list is the input to the next pass — without it, humanize has nothing to work with but the prose itself.

This skill is the agent-side input to Gate 4 (copy: true, in-voice, nothing fabricated). It produces the findings; the human makes the call.