Review Agent

Use the review stage to move drafts from “readable” to “ready for readers”.

When to use it

• After a chapter draft is done
• Before publication
• After multi-author edits
• To catch cross-chapter inconsistencies
• When you need read-only evidence paths before editing

Switch stage

Choose review in the AI input dropdown. History is shared with writing/planning in the same conversation.

Review checklist (suggested order)

1. Facts — numbers, names, dates, citations
2. Logic — argument flow, missing steps
3. Structure — headings, sections, redundancy
4. Prose — clarity, repetition, filler
5. “AI tone” — generic phrasing
6. Style — match @system rules
7. Format — headings, lists, terminology

Lightweight SOP

1. Read in chunks (chapter or line ranges)
2. Fix high/medium issues first with targeted edits
3. Re-scan for global consistency
4. Summarize: fixed items, remaining risks

Prefer edit_file with partial replacements; full rewrites only when most of the section must change.

For large review scopes, start with the review_project subagent. It scans read-only, returns evidence paths, and may save a reference report under @system/@subagent/; you can then decide what the review stage should actually edit.

Prompt templates

Full pass

You are in review mode. Review “{file}” for facts, logic, structure, tone, and format. List issues, fix high/medium items inline, then summarize changes and residual risks.

Minimal edits

You are in review mode. Make only necessary edits. Fix factual/logic errors first; do not rewrite voice unless required.

Style alignment

You are in review mode. Unify terminology and tone across {files} per @system writing style. Output a short change log.

Output shape (example)

File: {path}
Range: {start}-{end}

1) Issues
- [High] …
- [Med] …

2) Applied fixes
- …

3) Open risks
- …

If you need a project skeleton first, see Template creation.