/help

Autoresearch Help

Explain to the user:

Universal Autoresearch Plugin

An autonomous research loop for Claude Code. The agent iteratively modifies code, runs experiments, and a separate evaluation sub-agent judges results in a blind review.

Quick Start

/autoresearch "Optimize bundle size" --max-iterations 30

Or with a config file:

/autoresearch --config ./autoresearch.config.md

How It Works

1. Agent reads autoresearch.config.md for goal + evaluation criteria
2. Agent reads experiment-journal.md for history
3. Agent formulates hypothesis, modifies code, git commits
4. Agent runs experiment, collects results
5. Evaluation sub-agent (blind review): receives only evaluation criteria + results, decides keep/discard
6. keep = branch advances / discard = git reset
7. Journal + results.tsv updated
8. Loop continues via stop hook

Files Created in Your Project

File	Purpose	Git tracked?
autoresearch.config.md	Goal, evaluation criteria, constraints	Yes (commit)
experiment-journal.md	Structured experiment history	No
results.tsv	Quick-reference metrics log	No
run.log	Latest experiment output	No
.claude/autoresearch.local.md	Loop state (internal)	No

Commands

• /autoresearch [GOAL] [OPTIONS] — Start the loop
• /cancel-autoresearch — Stop the loop
• /autoresearch-help — This help

Options

• --config <path> — Config file path (default: ./autoresearch.config.md)
• --max-iterations <n> — Max experiments (default: unlimited)
• --max-consecutive-failures <n> — Stop after N failures (default: 5)