Run /iris-load at every session start and after every context condensation.
If /iris-load is unavailable, run the bash block from AGENTS.md.
Bias towards action. Default to action. Complete the job end-to-end. Never pause to ask permission for routine steps. Only prompt for destructive/dangerous operations.
Respect engineer availability. When the engineer is away, don't block on questions. Make reasonable decisions and push forward. Prefer conservative-but-complete over stopping halfway.
Escalate when stuck. After 2-3 failed attempts: reload ai.knowledge indexes
(/iris-load), consult Codex (/codex), go multimodal (screenshots, DOM dumps),
frame-shift. Stop guessing, start observing.
Completed job definition. A job is not done until: code implemented, tests written and passing, self-review against rules, PR opened with cross-repo links, CI monitored, review feedback addressed, comment threads resolved.
CLI tool path resolution. Never report a CLI tool as "not found" without checking:
Homebrew (/opt/homebrew/bin/), nvm (~/.nvm/versions/node/*/bin/), and npm global
prefix. Source nvm before running node/npm/codex commands:
export NVM_DIR="$HOME/.nvm" && [ -s "$NVM_DIR/nvm.sh" ] && . "$NVM_DIR/nvm.sh"
Full rules: ~/iris/ai.knowledge/rules/global/*.md
When the user's request matches a skill, invoke it using the Skill tool as your FIRST action. Do NOT answer directly first.
- Specs, "new feature", "spec this" →
/iris-spec - "Add a rule", agent mistake →
/iris-rule - "Document this", "the agent should know" →
/iris-learn - "Hand off to engineering" →
/iris-handoff - "Set up my machine", "onboard" →
/iris-onboard - "Sync", "submit rules" →
/iris-sync - "Install AGENTS.md", "install CLAUDE.md" →
/iris-install - "Audit", "review my changes" →
/iris-audit - "Clone", "get the repo" →
/iris-clone - "Run iris", "build from source" →
/iris-run - Bugs, errors, "why is this broken" →
/investigate - Ship, deploy, push, create PR →
/iris-ship - QA, test the site, find bugs →
/qa - Code review, check my diff →
/review - Performance, benchmarks →
/benchmark
Work in a branch. Only open a PR when asked. Never push to main.