review

Multi-Persona Code Review

Dispatch persona agents in parallel to review code from diverse expert perspectives. Each persona reviews independently through their unique philosophical lens, then findings are synthesized into a unified report.

Argument Parsing

Parse $ARGUMENTS to extract:

1. Review target: The non-flag portion — a file path, directory, or glob pattern. If empty, default to staged changes via git diff --staged.

2. --only filter: Comma-separated persona names (agent names or display names). Only these personas run.

3. --min-confidence N: Minimum confidence threshold for synthesis filtering (default: 30). Critical findings are never filtered.

Examples

$ARGUMENTS	Target	Personas	Min Confidence
src/auth.ts	src/auth.ts	All	30
--only "ThePrimeagen,DHH"	Staged changes	theprimeagen, dhh	30
src/api/ --only theprimeagen --min-confidence 60	src/api/	theprimeagen	60
(empty)	Staged changes	All	30

Agent Discovery

1. Glob ${CLAUDE_SKILL_DIR}/../../agents/*.md to find all persona files.
2. Exclude template.md.
3. Agent name = filename without .md (e.g., theprimeagen.md → theprimeagen).
4. Display name: check Persona Roster in ${CLAUDE_SKILL_DIR}/reference.md, or extract from # Claude Persona: {Name} heading, or convert kebab-case to title case.

--only Resolution

Match each name against discovered agents:

• Exact agent name match → use directly
• Display name match from Persona Roster → map to agent name
• No match → warn and skip

Pre-Dispatch

Show confirmation, then proceed immediately (no approval gate):

Personas: ThePrimeagen, DHH, Chris Coyier, ...
Target: src/auth.ts

Clean stale output:

rm -f persona-reviews/*.json persona-reviews/SYNTHESIS.md
mkdir -p persona-reviews

Dispatch

CRITICAL: Use the Agent tool to dispatch each persona. Do NOT review the code yourself. Do NOT skip delegation. Do NOT read agent files yourself — the Agent tool loads the persona automatically when you set subagent_type to the agent name.

For each selected persona, dispatch using the Agent tool:

• Set subagent_type to the persona's agent name (e.g., theprimeagen, dhh, chris-coyier)
• Set description to a short label like "ThePrimeagen reviews src/auth.ts"

The prompt to each persona should be simple and direct:

Review this code: {target}

{If file/directory/glob: "Read the files and review them through your lens."}
{If staged changes: include the git diff --staged output}

Give me your findings. For each issue you find, tell me:
- Severity (critical, warning, or suggestion)
- Confidence (0-100)
- File and line
- What's wrong
- What to do instead
- Why it matters — in YOUR voice

Be yourself. Write the way you actually talk.

Dispatch ALL personas in parallel. Issue all Agent tool calls in a single message. Do NOT dispatch sequentially.

Collection

As each persona finishes, note their completion. The Agent tool returns their findings directly — no JSON parsing needed, no files to write during collection.

Synthesis

After ALL personas complete:

1. Follow the Synthesis Protocol in ${CLAUDE_SKILL_DIR}/reference.md:
   • Deduplicate findings across personas (same file + semantically similar issue)
   • Boost confidence when multiple personas agree
   • Surface disagreements where personas conflict
   • Filter by --min-confidence threshold (critical findings never filtered)
2. Present the synthesized review to the user
3. Write synthesis to persona-reviews/SYNTHESIS.md