Doc Review

doc-review

Multi-agent review for markdown artifacts. Specs, plans, decision records, RFCs, design docs — anything written that another human (or another agent) is supposed to act on.

This skill wraps jutsu swarm doc-review. Three agents read the same artifact through three lenses, the orchestrator clusters their findings into a disagreement table, and a synthesizer agent writes the prose. Iterate via --replay while tuning prompts.

Why this exists

Before doc-review, kaijutsu's artifact-producing skills (spec-driven-development, planning-and-task-breakdown) each invented their own "fresh-eyes pass" mechanism. Each was different, none were systematic. This skill is the universal QA gate they all delegate to.

Layout (rich)

doc-review/
├── SKILL.md                     (this file)
├── skill.yaml
├── scripts/
│   └── run.sh                   thin wrapper around jutsu swarm doc-review
├── prompts/
│   ├── claude.md                completeness lens
│   ├── codex.md                 implementability lens
│   ├── antigravity.md                consistency lens
│   ├── synthesizer.md           cluster + render markdown
│   └── debate.md                Pass-2 critique template
├── references/
│   ├── lens-design.md           why each lens is what it is
│   └── interpreting-output.md   how to read the disagreement table
└── runbooks/
    └── tuning-prompts.md        per-repo override workflow

The CLI loads prompts from prompts/ at runtime and falls back to built-in defaults when a file is missing. Per-repo overrides go in /.claude/skills/doc-review/prompts/.md.

First-run consent

Before the first invocation in a repo, doc-review (like every swarm preset) requires explicit per-repo consent because input goes to remote model providers (Anthropic, OpenAI, Google).

If you run interactively from a TTY, the first invocation prompts and persists the answer. If you run headless (from inside an agent CLI session, from CI, or piped), the prompt has nothing to read from and the run aborts with a hint. Pre-grant consent with:

jutsu swarm doc-review --grant-consent

This writes allow-multi-model: true to .kaijutsu/doc-review.yaml and exits without running the swarm. Subsequent runs in the same repo work non-interactively.

When to invoke

Common entry points:

# After writing a spec
jutsu swarm doc-review SPEC.md

# Plan + decision review
jutsu swarm doc-review IMPLEMENTATION_PLAN.md
jutsu swarm doc-review .claude/decisions/2026-05-05-auth.md

# Multiple files (single concatenated review)
jutsu swarm doc-review SPEC.md PLAN.md

# --full for a high-stakes doc
jutsu swarm doc-review RFC.md --full --strict

# Iterate on the synthesis prompt
jutsu swarm doc-review --replay 

What each lens looks for

claude — completeness

• Missing edge cases the artifact should address but doesn't
• Undefined terms or ambiguous phrasing
• Scope creep (sections claiming work outside the stated goal)
• Internal contradictions across sections
• Acceptance criteria that aren't testable as written

codex — implementability

• "This section says X but never specifies HOW" — vague directives
• Acceptance criteria that don't say what passes vs fails
• Risks named without mitigations, or mitigations cited without the risk they address
• API/CLI/data-shape claims that contradict the rest of the document
• Numerical thresholds, timeouts, or limits left unquantified

antigravity — consistency

• Phrases or terms used differently in different sections
• References to other documents/sections/issues that don't resolve
• Contradictions between locked-decisions tables and stage-detail sections
• Drift from the document's own stated patterns
• Heading hierarchy gaps

Reading the output

Same structure as pr-review:

1. Disagreement table — rows are clustered findings, columns are agents. The 1/N rows (one agent flags, others didn't) are the conversation-starters. Read those first.
2. Synthesis — synthesizer's prose review.
3. Disagreements section — explicit callout of 1/N findings.
4. Per-agent stats footer — finding counts + costs.

See references/interpreting-output.md for the full rubric.

Cache + replay

Cache key is SHA256(preset + canonical body) → first 12 hex chars. Same artifact bytes → same cache → free --replay for tuning the synthesis prompt without re-spending tokens.

# First run: 3 agents review + synthesize, cache written
jutsu swarm doc-review SPEC.md
# stderr: ... cache key abc123def456 ...

# Edit synthesizer prompt, then replay (no model calls)
$EDITOR .claude/skills/doc-review/prompts/synthesizer.md
jutsu swarm doc-review --replay abc123def456

Workflow when invoked from another skill

spec-driven-development ends with: "Run jutsu swarm doc-review . Iterate findings until the disagreement table is mostly 2/N consensus or 0 findings."

planning-and-task-breakdown ends with: "Run jutsu swarm doc-review ."

The artifact-producing skill is responsible for surfacing the findings to the user and iterating. doc-review just runs the swarm.

Hard rules

• Never modify the artifact under review. doc-review is read-only by design — output goes to stdout, the user decides what to change.
• Never post comments. doc-review's input isn't tied to a PR; --post-comment is silently ignored even if pulled in by accident.
• Always cite section names in finding reasoning when possible. "Acceptance Criteria, paragraph 3" is more useful than "line 47".
• Don't double-review by running pr-review on the same markdown — pr-review's lenses are code-tuned, doc-review's are prose-tuned. Mixing them produces noise.

When NOT to use this

• Code reviews — use pr-review instead.
• Unstructured chat / brainstorm dumps — too noisy for the lenses. Tighten the artifact first.
• Live drafts — let the author finish a draft before reviewing. doc-review on incomplete writing produces lots of false-positive "missing" findings.

Source & license

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

• Author: momentmaker
• Source: momentmaker/kaijutsu
• License: MIT
• Homepage: https://kaijutsu.dev/

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