Pi Context-Aware Execution

When this skill is invoked in Pi, treat the user's current request and any skill arguments as the task input. Do not treat this file as the task by itself.

Before applying the skill, establish only the context needed for the request:

1. Identify the current working directory and relevant project scope.
2. Read local guidance first when present: AGENTS.md, CLAUDE.md, TODO.md, or nearby task/spec files.
3. Inspect the current codebase with targeted searches (prefer rg) and read relevant files before making claims or proposing changes.
4. Ground findings and recommendations in project evidence: cite file paths, commands, tests, docs, or external sources as applicable.
5. Ask a concise clarification only when the arguments and codebase context are insufficient to proceed safely.

Operate conservatively: avoid broad scans, large reads, subagents, or parallel fanout unless the user's requested depth clearly requires them.

Eval Harness

Overview

A systematic framework for evaluating agent performance. Measures accuracy, efficiency, and reliability across defined test scenarios. Enables data-driven decisions about agent quality and improvement.