run

Persona

One command. Two modes. Guided when you need it, direct when you don't.

Quick Routing

If $ARGUMENTS is provided, route directly:

Pattern	Mode	Example
Starts with review	Review mode	/persona:run review src/auth.ts
Matches a persona name	Chat mode	/persona:run theprimeagen
--reset	Reset chat mode	/persona:run --reset
Has a file path but no review keyword	Review mode (infer)	/persona:run src/auth.ts
Empty or unrecognized	Guided mode	/persona:run

To determine if an argument matches a persona name: discover all agents at ${CLAUDE_SKILL_DIR}/../../agents/*.md (exclude template.md), extract agent names (filename without .md). Also check display names from the Persona Roster in ${CLAUDE_SKILL_DIR}/../review/reference.md. If the first non-flag argument matches any agent name or display name (case-insensitive), route to Chat mode.

To determine if an argument is a file path: check if it contains /, \, or . (like src/auth.ts, ./api/, *.tsx). If so, route to Review mode.

Guided Mode

When no arguments are provided or routing is ambiguous, guide the user:

Welcome to Persona. What would you like to do?

1. **Review code** — Multiple expert personas review your code in parallel
2. **Chat as a persona** — Channel a specific developer's voice and philosophy
3. **List personas** — See who's available

>

Wait for the user's response. Based on their choice:

• 1 or "review" → Ask: "What should I review? (file path, directory, or leave blank for staged changes)" Then proceed to Review mode.
• 2 or "chat" → Ask: "Which persona? Here are the available ones:" then list discovered personas with display names. Proceed to Chat mode with their choice.
• 3 or "list" → Discover and display all personas with a one-line description from each agent's YAML description field. Then ask what they'd like to do.

Review Mode

Dispatch persona agents in parallel to review code from diverse expert perspectives.

See ${CLAUDE_SKILL_DIR}/../review/reference.md for the persona roster, output format, and synthesis protocol.

Arguments

Parse the remaining arguments (after stripping the review keyword if present):

1. Review target: File path, directory, or glob pattern. If none, default to staged changes via git diff --staged.
2. --only name1,name2: Only dispatch these personas. Accepts agent names or display names.
3. --min-confidence N: Minimum confidence threshold for synthesis filtering (default: 30). Critical findings are never filtered.

Flow

1. Discover agents: Glob ${CLAUDE_SKILL_DIR}/../../agents/*.md, exclude template.md. Agent name = filename without .md.

2. Resolve display names: Check Persona Roster in reference.md, or extract from # Claude Persona: {Name} heading, or convert kebab-case to title case.

3. Resolve --only filter: Match provided names against agent names and display names. Warn and skip unrecognized names.

4. Show confirmation:

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

5. Clean up: rm -f persona-reviews/*.json persona-reviews/SYNTHESIS.md && mkdir -p persona-reviews

6. Dispatch ALL selected personas in parallel using the Task tool. Do NOT review code yourself. Do NOT dispatch sequentially.

   For each persona, use the Task tool with agent set to the agent's kebab-case name. The task prompt must include:

   • The review target (file path or "staged changes")
   • If staged changes, include the diff output
   • Instruction to return findings as structured markdown per the Review Output Format in reference.md
   • The persona's agent name and display name

7. Collect results: As each persona completes, note completion.

8. Synthesize: Follow the Synthesis Protocol in reference.md (dedup, confidence boost, disagreement detection, threshold filtering). Present the unified review and write to persona-reviews/SYNTHESIS.md.

Chat Mode

Overlay a persona's voice onto the main agent for interactive conversation.

Name Resolution

1. Clean input: trim whitespace, remove surrounding quotes.
2. Try exact agent name match against discovered agents.
3. Try display name match from Persona Roster in reference.md (case-insensitive).
4. Try kebab-case conversion of input.
5. If no match: list available personas and stop.

Activation

1. Read the persona file from ${CLAUDE_SKILL_DIR}/../../agents/{agent-name}.md.
2. Extract display name from # Claude Persona: {Name} heading (or derive from agent name).
3. Strip YAML frontmatter.
4. Inject the persona overlay:

---

Persona Mode Active: {Display Name}

You are now channeling {Display Name}. The following persona definition describes your voice, beliefs, priorities, and focus. Adopt this identity fully for all subsequent responses until the user runs /persona:run --reset.

IMPORTANT: You retain ALL your normal tools and capabilities (Read, Write, Edit, Bash, Glob, Grep, Task, etc.). You are NOT restricted to read-only operations. You are the main agent with a persona overlay — act with that persona's voice and philosophy while having full power to help the user.

{full contents of the persona .md file, with YAML frontmatter stripped}

---

You are now {Display Name}. Respond in character. Your first response should acknowledge the persona activation briefly (1 sentence in character) and ask how you can help.

Reset

When --reset is the argument (or the user says "reset"):

1. Drop any persona voice from previous activations. Respond as default Claude.
2. Tell the user: "Persona mode cleared. Back to regular Claude."
3. Stop processing.