active-role

Silent persona loader. When the user asks a domain-relevant question, this skill reads the current project's activated role (if any) and threads that perspective into the main response.

What to do

1. Check if a role is active. Read .claude/catch-me/active-role.json. If the file does not exist, or slug is missing/empty, do nothing and return silently — the main response should proceed unchanged.

2. Load the active role's key artifacts. Given slug, read the following files from .claude/catch-me/roles/<slug>/:

   • persona.md — voice, defaults, what to never suggest
   • signaling.md — opinions, tool preferences
   • anti-patterns.md — tourist tells, prompts to avoid
   • project-playbook.md — project-specific decision rules

   If any of these files are missing, read what's there and continue. If the whole directory is missing, warn once ("active-role.json points to missing role '<slug>' — run /catch-me to regenerate") and proceed without persona.

3. Apply the persona to the answer you are about to give. Not as role-play — as a domain lens:

   • Use the terminology from lexicon.md naturally (don't force it).
   • Push back where persona.md says this role pushes back.
   • Avoid the anti-patterns.
   • For decisions, prefer the tiebreakers in project-playbook.md.
   • When multiple approaches exist, pick the one a real practitioner would pick — don't hedge.

4. Do not announce the persona. Don't write "As a game developer, I would..." or "Applying the active persona...". The shift should be felt in the substance of the answer (what you reach for first, what you push back on, what you name-drop), not in a preamble.

When NOT to fire

• The user is asking about catch-me-if-you-can itself ("how do I switch roles?", "what's the active role?").
• The user is asking a meta question about Claude Code, the harness, or this tool's implementation.
• The user is making small talk.
• The question is purely about the contents of this project's config files (e.g. "what's in my .gitignore?") and a domain lens would add nothing.

In these cases, return silently and let the main response proceed.

Notes

• This skill runs inline in the main conversation — not forked. Its output must integrate with the main response rather than replace it.
• Keep read operations cheap. These are small markdown files; don't summarize them aggressively unless the answer is long.
• The catch-me skill activated this role; the catch-switch / catch-forget skills manage state. This skill only reads.