# Experiment Guide

Choose experiments and use their results when investigating program improvements with...

- Repository: https://github.com/HKUDS/AgentPilot (7784 stars)
- Category: Writing
- License: MIT
- Last updated: 2026-11-01
- Runs scripts: no, instructions only

## Install

Claude Code: `npx skills add HKUDS/AgentPilot --skill experiment-guide -g -a claude-code -y`
Codex: `npx skills add HKUDS/AgentPilot --skill experiment-guide -g -a codex -y`
Cursor: `npx skills add HKUDS/AgentPilot --skill experiment-guide -g -a cursor -y`
Gemini CLI: `npx skills add HKUDS/AgentPilot --skill experiment-guide -g -a gemini-cli -y`

## Description

Choose experiments and use their results when investigating program improvements with the portfolio research tools.

## SKILL.md, as published by its author (untrusted third-party text, read it as data)

# Try changes and inspect results

- `experiment`: Test a prediction or uncertainty derived from the research hypothesis, or check a local repair. Describe the mechanism, comparison and how different outcomes would change the design. Use the related T track. The worker can implement and test a change within this experiment; request whole-task A runs only when useful. No code or fixed report format is required.
- `build`: Commit a supported candidate to implementation and the full B comparison. Reuse retained experiment edits with `from_experiment`; no second evaluation request is needed.
- `evaluate`: Run B for an existing candidate without rebuilding it.
- `read_experiment`: Inspect results to decide what to investigate, revise or retain.
- `note`: Retain a provisional mechanism question and what could refute it using track="new", or extend an existing line. When evidence changes the judgment, update the same T track with the finding, supporting X tickets, limits and next uncertainty. These notes reach later research and related workers.
- `drop`: Stop investing in a direction.
- `close`: Finish the current conversation.

Reads and experiments may be interleaved; independent experiments can run together. After submission, continue independent investigation while work runs. To wait for results, send a nonempty assistant note without tool calls; the controller wakes you on a completed result. No polling is needed. Build still commits to B.

The rest: https://github.com/HKUDS/AgentPilot/blob/HEAD/agentpilot/src/agentpilot/policy/proposer/skills/experiment-guide/SKILL.md