/autoresearch
Start, resume, inspect, report, or pause a measured experiment loop
$ npx -y skills add drivelineresearch/autoresearch-claude-code --agent claude-codeHow it fires
How this command gets triggered: by you, by Claude, or both.
- Fires itselfClaude auto-loads it when your prompt matches the work.
- You can call itInvoke it directly when you want it.
- Slash command
/autoresearch
Context preview
What this command does when you run it.
Start, resume, inspect, report, or pause a measured experiment loop
Command definition
autoresearch.mddescription: Start, resume, inspect, report, or pause a measured experiment loop
argument-hint: [off | status | report | goal description]
allowed-tools:
- Read
- Write
- Edit
- Bash
- Glob
- Grep
Autoresearch command
Arguments: $ARGUMENTS
Read `${CLAUDE_PLUGIN_ROOT}/skills/autoresearch/SKILL.md` and follow its protocol. Claude substitutes that plugin path in this command's text; it need not exist as an environment variable in Bash. For a manual install, where the placeholder is not expanded, read `~/.claude/skills/autoresearch/SKILL.md` instead. Resolve `AR_SCRIPTS` to the `scripts` directory beside the file you read. Do not invoke `autoresearch` through the Skill tool again: this adapter shares that name and can shadow the shared skill. If neither known path exists, pause and report the missing installation rather than searching outside those locations.
Handle the requested mode before any resume or setup action:
- **off:** create `.autoresearch-off` in the experiment workspace, preserve partial
work, and stop. Do not launch another experiment.
- **status:** read state/dashboard/recent worklog and show current segment budget,
baseline/best, counts, and pause status. Do not resume or change experiment data. For a live Claude session, first create `.autoresearch-inspect` so the Stop hook permits this inspection turn to end. This small control write is the only write.
- **report:** for a live Claude session create `.autoresearch-inspect`; write
`autoresearch-report.md` with results, evidence limits, winning changes, failures, and remaining ideas. Do not run experiments or unpause. This mode writes a report.
- **resume (session exists):** read session/state/worklog and git status/log. Resolve
partial/unlogged work first and check the budget. Remove `.autoresearch-off` only when explicitly resuming within budget; do not reset the config to gain runs. An exhausted budget needs a user extension or deliberate new experiment contract.
- **fresh goal (no session):** verify git and preserve existing work, then follow
skill setup. Infer goal/metric/scope from supplied arguments where possible.
A user stop, cancellation, or new scope overrides continuation. Complete one experiment per autonomous turn; the Claude Stop hook handles the next turn when valid state and budget permit it. Never treat a plugin code review as a request to start an optimization session.
Read more
description: Start, resume, inspect, report, or pause a measured experiment loop argument-hint: [off | status | report | goal description] allowed-tools: - Read - Write - Edit - Bash - Glob - Grep
Autoresearch command
Arguments: $ARGUMENTS
Read `${CLAUDE_PLUGIN_ROOT}/skills/autoresearch/SKILL.md` and follow its protocol. Claude substitutes that plugin path in this command's text; it need not exist as an environment variable in Bash. For a manual install, where the placeholder is not expanded, read `~/.claude/skills/autoresearch/SKILL.md` instead. Resolve `AR_SCRIPTS` to the `scripts` directory beside the file you read. Do not invoke `autoresearch` through the Skill tool again: this adapter shares that name and can shadow the shared skill. If neither known path exists, pause and report the missing installation rather than searching outside those locations.
Handle the requested mode before any resume or setup action:
- **off:** create `.autoresearch-off` in the experiment workspace, preserve partial
work, and stop. Do not launch another experiment.
- **status:** read state/dashboard/recent worklog and show current segment budget,
baseline/best, counts, and pause status. Do not resume or change experiment data. For a live Claude session, first create `.autoresearch-inspect` so the Stop hook permits this inspection turn to end. This small control write is the only write.
- **report:** for a live Claude session create `.autoresearch-inspect`; write
`autoresearch-report.md` with results, evidence limits, winning changes, failures, and remaining ideas. Do not run experiments or unpause. This mode writes a report.
- **resume (session exists):** read session/state/worklog and git status/log. Resolve
partial/unlogged work first and check the budget. Remove `.autoresearch-off` only when explicitly resuming within budget; do not reset the config to gain runs. An exhausted budget needs a user extension or deliberate new experiment contract.
- **fresh goal (no session):** verify git and preserve existing work, then follow
skill setup. Infer goal/metric/scope from supplied arguments where possible.
A user stop, cancellation, or new scope overrides continuation. Complete one experiment per autonomous turn; the Claude Stop hook handles the next turn when valid state and budget permit it. Never treat a plugin code review as a request to start an optimization session.
A measured experiment loop for Claude Code and Codex. Define a goal, a fixed benchmark, and files to modify. The agent tries changes, measures results, commits supported improvements, and records failed ideas within a budget. Inspired by pi-autoresearch.