Stats
Actions
Tags
From claude-pi
Use when running a superpowers workflow — subagent-driven-development, dispatching-parallel-agents, systematic-debugging, or requesting-code-review — to decide which roles should go to Pi versus the Agent tool
How this skill is triggered — by the user, by Claude, or both
Slash command
/claude-pi:superpowersThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Pi provides a heterogeneous agent fleet — Gemini, GPT-4, Claude, and others — running outside your context window. Superpowers workflows have natural read-only roles where Pi adds model diversity and breadth without consuming tokens.
Pi provides a heterogeneous agent fleet — Gemini, GPT-4, Claude, and others — running outside your context window. Superpowers workflows have natural read-only roles where Pi adds model diversity and breadth without consuming tokens.
REQUIRED: Load claude-pi:orchestrate for Pi tool signatures, model tiers, and the parallel pattern.
| Role | Tool | Tier |
|---|---|---|
| Spec compliance reviewer | Pi | balanced |
| Code quality reviewer | Pi | balanced |
| Security reviewer | Pi | frontier |
| Parallel investigation agent | Pi | balanced |
| Parallel research agent | Pi | fast |
| Implementer subagent | Agent tool | — needs Bash/Edit/Write |
| Brainstorming dialogue | neither | — no subagent dispatch |
Reviewer roles receive the full diff + spec context in the prompt — no local file access needed.
Per task:
Implementer → Agent tool (needs Edit/Bash)
Spec reviewer → pi_spawn_agent balanced
Code quality reviewer → pi_spawn_agent balanced
Security-sensitive → pi_spawn_agent frontier
Use pi_run_task for sequential reviews. For the final cross-task review, spawn once and pi_get_result.
Use followUp: true on pi_spawn_agent when you want to queue a review without interrupting the implementer.
Independent investigations are Pi's strongest use case — each agent gets a self-contained prompt.
1. pi_spawn_agent × N — focused investigation prompt per agent
2. pi_poll_agent loop — steer if off-track
3. pi_get_result × N — collect findings
4. synthesize — combine in this session
Prefer balanced for code/log analysis; fast for pure text retrieval.
The reviewer receives git SHAs and a full diff — entirely read-only and ideal for Pi.
pi_run_task(
model: 'google/gemini-2.5-pro', // frontier when review gates a merge to main
task: <full requesting-code-review prompt with git diff>
)
Use balanced for draft or in-progress reviews; frontier when gating main.
Pi agents have no access to your local filesystem. Never delegate roles that require:
When a role needs local access, use the Agent tool.
npx claudepluginhub crichalchemist/agent-pi --plugin claude-piCreates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.