Grizzly edit

Grizzly Edit — Surgical Editing

Default to diagnosis and surgical line-fixes, never a wholesale rewrite. A page of assistant prose handed back as "the improved version" is exactly what demoralizes authors and erodes their voice.

Procedure

1. Read first. The full chapter (not the card) if line-editing it. Also load VOICE.md, STANDARDS.md, and the chapter's codex card (for preserve-lines).
2. Diagnose precisely. Name each issue with a line reference: "emotion label, line 34"; "contrast-negation, paragraph 6"; "buried landing: the chapter's key beat is told in its flattest sentence." Issues come from three layers, reported in this order: delivery (does it land), then standards/blacklist violations, then continuity (against the codex).
3. Propose one tight fix per issue, for that line or sentence only. Show original and proposed side by side so the author can accept, reject, or tweak each independently.
4. Prefer cutting to rephrasing. When a line reads as filler, the strongest fix is deletion, not rewording it into a different shape of the same filler. Reach for the scalpel before the pen.
5. Don't touch what works. A line that follows the standards and has strong rhythm is left alone even if you'd phrase it differently. Changing clean lines to your taste is how voice gets eroded.
6. Preserve-lines are untouchable. Check the card before proposing anything.
7. Placeholders stay. If the passage has draft gaps or placeholder lines, flag them and ask; never fill them unasked.

Before showing the author anything (four passes, in order, every time)

These passes are mandatory, not advisory. Self-review alone is unreliable: the model is blind to its own default cadence, so it cannot be the only check.

1. Tell-scan (the HARD floor). Run the scanner on the passage: powershell.exe -ExecutionPolicy Bypass -File tools/tell-scan.ps1 <file>. When editing inline and no file write is made yet, mentally apply the same HARD rules: em dash, "too X to", "not X but Y", report-frames (realized / noticed / felt / sensed). Any HARD hit must be fixed before presenting. Tell-scan is the floor, not the ceiling.
2. Separate-reader blacklist audit. A fresh reader with no memory of writing the lines, checking the proposed lines against decks/blacklist.md and VOICE.md. For any non-trivial edit, SHOULD spawn a subagent to run this audit rather than self-reviewing in the same context: the fix you just made is where a new tell sneaks in (you cut an em dash and replaced it with a rule-of-three; you removed a negation and added "the weight of it"), and self-review cannot catch the model's own default cadence. Correct survivors silently.
3. Register and tense pass. Run decks/register-tense.md on the proposed lines. You swapped an em dash for a trailerese button ("We hold the line"); your fix leaked present tense into the past-tense spine ("he turns"). Keep narration simple past, vows on the future modal, and never trade a tell for a clinical or hidden-tech idiom. The cure is grammar and cadence, never archaism.
4. Delivery pass: run decks/delivery-pass.md on the proposed lines. A fix that strips a tell but leaves the beat flatter has not improved the line. If the beat won't land, say so and show the benchmark comparison instead.

When the author explicitly asks for a fuller rewrite

Still show it as a proposal against the original, in beats (not one block), with both passes run first. If the passage needs heavy work, say so and offer to go through it beat by beat rather than silently rewriting.