Gen Brief

Generate revision brief

A revision brief is a focused instruction to the writer, NOT a full re-spec of the chapter. It says: keep what works, change THIS specifically, here's what to add or cut, here are the constraints.

The brief replaces hand-written editorial notes. Quality of the brief drives quality of the revision.

Inputs

Caller passes either:

• --auto — pick the chapter automatically based on lowest score / strongest consensus signal
• --chapter N — a specific chapter

For each chapter under consideration, gather (where present):

• /chapters/ch_.md
• Latest /eval_logs/*_chNN.json for that chapter
• Latest /eval_logs/*_full.json and any top_suggestion/weakest_chapter references
• /edit_logs/reader_panel.json — extract any consensus items mentioning this chapter
• /edit_logs/chNN_cuts.json — adversarial cut data
• Latest /reviews/*_parsed.json — Opus items mentioning this chapter

Auto-pick logic

If --auto:

1. From reader_panel.json consensus items: any cut_candidate, missing_scene, or worst_scene with 3-of-4 or 4-of-4 agreement → pick that chapter, brief shape inferred from the question.
2. If no consensus, pick the chapter with lowest overall_score in latest eval logs.
3. If scores are tied within 0.5 across the bottom three, pick the one with most adversarial cuts ≥17% fat.

Brief shapes

Pick the shape that fits the dominant signal:

A. COMPRESSION (cut_candidate consensus, or chapter > target word count by >20%)

Target: 40-60% word reduction. Identify the 2-3 essential beats (what the reader panel said worked or what the outline marks as load-bearing). Cut everything else, but DO NOT compress below 1800 words — sweet spot is 2200-3000 for a compressed chapter.

B. EXPANSION (missing_scene consensus, or chapter is too short for its structural importance)

Target: +800-1500 words. Specify WHAT BEATS to expand (the missing scene from panel feedback, or the under-developed beat from eval). Specify WHAT TO KEEP unchanged. Don't write "make it longer" — write "add a private moment between X and Y after the existing scene at the bell tower."

C. DRAMATISATION (weak scene, low engagement, summary-heavy)

Change HOW information arrives, not WHAT information. "Reading a document" → investigation/confrontation. "Briefing" → confrontation with resistance. The plot beats stay; the medium changes.

D. CHARACTER DEEPENING (thin_character consensus)

Identify 1-2 existing scenes where the character appears. Add a private/unguarded moment the POV catches. Connect to backstory in characters.md. Do NOT add a new scene — deepen an existing one.

E. CONSISTENCY / TIMELINE FIX (canoncompliance violations, contradictionsfound)

Specify the contradiction. Specify which value is canonical. Specify all places that need to update (canon.md + outline.md + the affected chapters).

F. VOICE / SLOP CLEANUP (high slop_penalty, voice fingerprint outlier)

Bullet-list specific patterns to fix: "remove all 7 instances of 'a sense of'", "reduce em-dash density from 23/1k to ≤15/1k", "vary sentence-start: only 18 of 47 sentences should start with 'He' — currently 31."

Brief format (write to /briefs/ch_.md)

# Revision brief — Ch  () — shape: 

## Why this revision

## What to keep

## What to change

## Constraints

- Word count target:  (range, with reason)
- Voice: stay within voice.md Part 2; in particular, watch for 
- Continuity: ending must flow into Ch 's opening (currently: "")
- Canon: do not contradict 

## Anti-patterns to avoid (from prior versions)

Output

Write the brief, then print:

• The chosen chapter and brief shape
• A one-line summary of the change
• The expected word count delta

Suggest: gen-revision briefs/ch_.md.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: tchr-dev
• Source: tchr-dev/autonovel
• License: MIT

Install and usage instructions live in the source repository linked above.