pi-delegate_skill

Overview

This skill lets you run Pi as a non-interactive subagent to handle bounded tasks with fresh context, a different model, or isolated execution. It provides command patterns, model recommendations, and prompt guidance so you can delegate work reliably from a project directory. Use it to get an independent pass, critique, or noise-isolated processing without changing the main conversation state.

How this skill works

The skill runs Pi in a separate process with fixed flags to avoid interactive behaviors and shared templates. It enforces changing to the target working directory, adding offline and no-template flags, and optionally selecting a model or continuing a prior delegated task. Output is returned as a single response containing the requested findings, recommendations, or next actions.

When to use it

• Delegate a bounded, self-contained task away from the main conversation
• Run a different model for specialized capabilities or cost trade-offs
• Get an independent second opinion or critique on a plan or document
• Isolate noisy or exploratory work so it doesn't contaminate current context
• Repeat or continue a previous delegated job with -c/--continue

Best practices

• cd into the target project directory before running Pi to scope files and context
• Always include --offline --no-prompt-templates --no-themes to ensure deterministic, non-interactive output
• Specify --model {model} when model behavior or capability matters
• Use -c/--continue to refine or extend the same delegated task across runs
• Keep prompts direct and request concrete outputs (findings, risks, options, recommendation, next actions)

Example use cases

• Review a proposal and return top risks plus recommended fixes using a capable model
• Ask for a second opinion on a design decision with a balanced or cheaper model for quick iteration
• Run a focused code or doc audit in the project directory and return a prioritized checklist
• Isolate experimental or noisy analysis so the main agent state stays clean
• Continue refining a recommendation by rerunning with -c to produce an updated plan