autoresearch-program

Autoresearch Program

Stamp out the loop directive (program.md) for an autoresearch run. This skill substitutes repo-specific values into a template and writes the result to the repo root. Once program.md exists, the autoresearch skills exit the picture — the human hands program.md to a fresh agent session and the loop runs from there.

program.md is voiced as a direct address to the loop agent: imperative, concrete, researcher-tone (LOOP FOREVER, What you CAN do / What you CANNOT do, NEVER STOP, Simplicity criterion, Rewinding). This voice is load-bearing — it sets the autonomy expectation and the discipline. Preserve it during substitution.

Prerequisites

• autoresearch-verify has been run and the verifier script exists. Confirm both modes work (light and heavy) by reading the verifier and, if possible, running each mode once.
• autoresearch-method has been read; repo fit is confirmed.
• The repo is a git repo. The working tree should be clean — program.md is a tracked file the team shares.

What this skill produces

• <repo root>/program.md — generated by substituting placeholders in ./program.template.md.
• .gitignore updates for the results-log path and any verifier persistence directory, if not already ignored.

The skill does not start the loop, create branches, or run the verifier.

Files in this skill

• ./program.template.md — the loop directive with {{PLACEHOLDER}} tokens. Read this file as the substitution source.

Information to gather

Read the repo and ask the human as needed. Settle every value before substituting.

• Mutable scope. Files or globs the loop agent is allowed to edit. Format as Markdown bullets. The narrower, the better — wider scope = noisier loop.
• Immutable scope. Files or globs the loop agent must not edit. Format as Markdown bullets. At minimum include: the verifier script, the verifier's persistence directory, tests/benchmarks/fixtures/datasets, lockfiles, CI files, and program.md itself.
• Branch prefix. Default: autoresearch. Each run creates <prefix>/<tag>.
• Light verifier command. Default: python verify.py --mode light. Override if the verify skill used a different filename or invocation.
• Heavy verifier command. Default: python verify.py --mode heavy. Same caveat.
• Results log path. Default: results.jsonl at the repo root.
• Acceptance rule. Default text: "Keep the candidate when light.valid is true AND light.score is strictly better than the current best (in the direction given by light.objective)." Adjust if the metric is noisy (require N consecutive better runs, require a margin above the noise floor, require a minimum sample size). The heavy gate is enforced separately by the loop and does not need to be repeated in this rule. Phrase the rule in plain English directed at the loop agent — it will read this prose, not parse it.

Placeholder map

Placeholder	Replaced by
{{MUTABLE_SCOPE}}	Mutable scope bullets
{{IMMUTABLE_SCOPE}}	Immutable scope bullets
{{BRANCH_PREFIX}}	Branch prefix
{{VERIFY_CMD_LIGHT}}	Light verifier command
{{VERIFY_CMD_HEAVY}}	Heavy verifier command
{{RESULTS_LOG}}	Results log path
{{ACCEPTANCE_RULE}}	Acceptance rule prose

Steps

1. Read the template

Read ./program.template.md from this skill directory.

2. Substitute placeholders

Perform a literal string replacement for each {{PLACEHOLDER}} listed above. Do not modify any other text. Specifically, leave verbatim:

• The experiment loop (LOOP FOREVER + numbered steps + advance-the-branch framing).
• What you CAN do / What you CANNOT do.
• Crashes handling.
• Simplicity criterion.
• Rewinding.
• Results Log schema.
• NEVER STOP with the sleep example.

These sections are the methodology. They carry the voice and the autonomy directive that make the loop work. If you find yourself wanting to rewrite a section, soften the tone, or add a new section, stop and surface the disagreement to the human.

3. Write program.md

Write the substituted content to <repo root>/program.md. If a program.md already exists, show the human the diff and confirm before overwriting.

4. Update .gitignore

If the results log path is not already gitignored, append it. Also ignore any verifier persistence directory referenced by the loop (e.g. verify_runs/). Do not gitignore the verifier script itself — it must be tracked.

5. Report and hand off

Tell the human:

• Where program.md was written.

• The mutable and immutable scope at a glance.

• The verifier commands baked in.

• The next step: start a fresh agent session and prompt it with:

  Read program.md and follow it exactly. Begin with setup, then enter the loop and do not stop until I interrupt you.

Boundaries

• Substitution only. Do not rewrite, summarize, soften, or extend the template's methodology sections (the experiment loop, What you CAN/CANNOT do, Crashes, Simplicity criterion, Rewinding, Results Log, NEVER STOP).
• Do not start the loop yourself. This skill produces a file; the loop runs from a fresh agent session reading program.md.
• Do not edit the verifier. If the verifier is wrong, return to autoresearch-verify.

After this skill

The autoresearch skills are no longer in the picture. The loop is self-contained:

• program.md is the operating directive.
• The verifier (verify.py or equivalent) produces light scores and heavy integrity verdicts.
• The results log is the durable record across runs.

To run, hand program.md to a fresh agent session.