Automake

Automake

Automake creates a git-backed improvement loop: build one focused candidate, commit it, judge it against the previous kept version, keep BETTER, and revert NOT_BETTER.

The keep rule is invariant: BETTER means the candidate is closer than the baseline to the automake prompt and judge criteria. If any generated prompt or judge wording conflicts with that rule, this skill wins.

Load References

Load only the relevant files:

• Current-agent default: references/current-agent-program.md
• Optional headless/script mode: references/headless-script.md
• Judge wording or subagent/headless judge setup: references/judge-contract.md
• Multiple judge roles, personas, or perspectives: references/panel-judging.md
• Run folder files, logging, and storage layout: references/run-folder-layout.md
• Iteration limits or subjective/adversarial judge tuning: references/tuning.md

Default Mode

Use current-agent mode unless the user explicitly asks for script, headless, shell, or automake.sh.

Current-agent mode writes only:

~/.automake//program.md

Headless-script mode writes only:

~/.automake//automake.sh
~/.automake//prompt.md
~/.automake//judge.md

Never write those files into the target repo.

Interview

Begin by interviewing the user. Do not create files until all required answers are collected.

Ask only missing direct inputs, one at a time:

1. What is your goal for this loop? This becomes the automake prompt.
2. What should the judge use to reject or accept candidates?

For current-agent mode, also ask:

3. Should the judge be a subagent or a headless agent?

Use defaults unless the user specifies otherwise:

max_iterations=5
max_consecutive_failures=3

Infer judge_strategy=panel without asking when the judge criteria clearly name multiple judge roles, personas, or perspectives. Examples: "technical judge and user judge", "one judge for implementation, one for UX, one for research", or a bullet list of distinct judges. Otherwise use judge_strategy=single.

Do not ask discovery, quality, or extra configuration questions. If the user provides multiple answers up front, parse and reuse them. Skip any question whose answer is already available from the request, previous turns, or the defaults above.

Use AskUserQuestion, Question, request_user_input, or the closest available structured-question tool when available. If no structured-question tool is available, ask one concise question in chat and wait for the answer.

Core Contract

• Require a committed, clean target repo baseline before running. The loop may use git reset --hard and git clean -fd in the target repo after user approval.
• Store every automake artifact under RUN_DIR=~/.automake//.
• Do not add .automake/, program.md, automake.sh, prompt.md, judge.md, results, judgments, logs, or run metadata to the target repo.
• Append durable results. Do not overwrite previous run history inside the run folder.
• The builder reads $RUN_DIR/results.tsv, $RUN_DIR/learnings.md, and recent $RUN_DIR/runs/*/judgments/*.md to avoid repeating failures.
• The judge gets fresh context. Do not pass prior judgments or builder logs as persuasive evidence.
• For panel judging, each panel judge gets independent fresh context and returns its own verdict before aggregation.
• REQUIRED FREEZE: while the judge is running, the main agent or script must do no work except polling for judge completion.
• The judge may inspect the repo and run checks, but must not mutate the candidate. If it does, reject and clean the mutation.
• Judge output must include evidence, not just rationale.
• Panel final verdict is BETTER only if every required panel judge returns BETTER; otherwise it is NOT_BETTER.
• Keep every candidate judged BETTER; revert every candidate judged NOT_BETTER.
• Stop when iteration >= max_iterations or failures >= max_consecutive_failures.

For exact judge wording, load references/judge-contract.md. For panel judging, load references/panel-judging.md.

Setup Workflow

After answers are complete:

1. Generate a three-random-word lowercase slug.
2. Create ~/.automake//.
3. Set RUN_DIR to that absolute run folder path.
4. Set REPO_DIR to the absolute path of the current working repo.
5. Verify REPO_DIR is a git repo.
6. Verify the target repo has a clean committed baseline, or ask the user to commit/stash before running.
7. Load references/run-folder-layout.md and initialize the required run folder scaffolding.
8. For current-agent mode, load references/current-agent-program.md and write program.md.
9. For headless-script mode only, load references/headless-script.md and write automake.sh, prompt.md, and judge.md.
10. Ask the user for explicit approval before running.

When final CLI flags are uncertain, inspect local help before writing commands:

claude --help
codex exec --help
cursor --help
opencode --help

Preserve the contract even when flags differ: non-interactive operation for headless calls, full tool access, no sandbox, skipped approvals, and fresh judge context.

Completion Checklist

Setup is complete only when:

• RUN_DIR exists under ~/.automake//.
• All generated automake files exist in RUN_DIR, not REPO_DIR.
• REPO_DIR is absolute and points at the target git repo.
• The target repo is clean and committed, or the user has been told exactly what must be committed/stashed before running.
• The chosen mode matches the user's request: current-agent by default, headless-script only when explicit.
• The chosen judge strategy matches the judge criteria: single by default, panel when multiple roles or perspectives are explicit.
• The generated files contain the prompt, judge criteria, iteration limits, consecutive failure limit, and keep/revert rule.
• The user has been asked whether to run now.

Ask for approval with this shape:

All done - ready to run?
1. Yes, run automake now
2. No, I want to change something first

If the user says yes:

• Current-agent mode: read and execute ~/.automake//program.md.
• Headless-script mode: run ~/.automake//automake.sh.

If the user says no, ask what they want changed, update the files in RUN_DIR, then ask again.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: eliasstravik
• Source: eliasstravik/skills
• License: MIT

Install and usage instructions live in the source repository linked above.