Back to search
autoloop
SkillWorks with Claude CodeGitHub
Use when the user types /autoloop, or asks for a spec-to-code run with no check-ins ("loop without checkins", "take all recommendations", "fable review the spec/plan/code", "until code complete, tested and reviewed"). Also use to resume when a status memory says AUTOLOOP RUNNING.
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Badge for your README
Markdown
[](/listings/gregprouty--autoloop-skill)HTML
<a href="/listings/gregprouty--autoloop-skill"><img src="/api/listings/gregprouty--autoloop-skill/badge.svg" alt="findsafeskills safety rating" /></a>Install
i
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
git clone https://github.com/GregProuty/autoloop-skill ~/.claude/skills/autoloop
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
SKILL.md · Manifest · 58 lines
- ---
- name: autoloop
- description: Use when the user types /autoloop, or asks for a spec-to-code run with no check-ins ("loop without checkins", "take all recommendations", "fable review the spec/plan/code", "until code complete, tested and reviewed"). Also use to resume when a status memory says AUTOLOOP RUNNING.
- ---
- # Autoloop
- The owner has walked away. Drive one piece of work to code that is complete, tested and reviewed. Never end a turn with a question or a menu.
- ## Owner's instruction: every sub-skill gate is pre-answered
- Any approval, consent, "which option?" or "ask your human partner" step in a superpowers skill is already answered: take the recommended option, record it as `(rec.)`, continue.
- | Gate | Answer |
- |---|---|
- | brainstorming: questions, section approvals, user spec review, visual companion | Recommended; approved; never offer the companion. Next is the Fable review, not writing-plans |
- | "Too big, decompose" | Split it yourself; build the parts in order |
- | using-git-worktrees: consent, native tool, failing baseline | `git worktree add .worktrees/<topic>`, set up per the project's worktree memories; fix or record baseline failures |
- | writing-plans: execution choice | Subagent-driven (inline for tier S) |
- | subagent-driven-development: pre-flight question, plan-mandated finding, "escalate" | Decide for the spec's intent, amend the plan, log it in the ledger |
- | subagent-driven-development: final review, then finishing | That review is step 6. Never invoke finishing-a-development-branch |
- | systematic-debugging: "discuss" after 3 failed fixes | Get a Fable root-cause diagnosis, continue |
- ## State
- First, write project memory `<topic>-status.md`, with a MEMORY.md line starting `AUTOLOOP RUNNING`:
- Order (verbatim): "<the user's message>"
- Tier: S|M|L Finish: spike|merge|branch Push: no
- Branch, worktree, base SHA, spec, plan, ledger, review files
- Done: <steps> Next: <step>
- Verified: <SHA, test count>
- Update `Done`/`Next` with every commit. After compaction: re-read this skill, that memory, `git log` and the SDD ledger; resume at `Next`; never redo a `Done` step.
- ## Tier (pick once, record, never downgrade)
- - **S**, about 3 source files, no new system or format: spec and task list in one doc, one Fable review of it, build inline with TDD, Fable code review.
- - **M**, one feature, up to about 8 tasks: every step.
- - **L**, bigger: split into parts, steps 2 to 6 per part; split the code review by area across parallel Fable reviewers.
- ## Steps
- 1. **Isolate.** `spike/<topic>` if called a spike, else a feature branch, from master. Run the baseline suite in the background.
- 2. **Spec.** superpowers:brainstorming. Commit.
- 3. **Fable review the spec** with `review-prompt.md` (this folder). Meanwhile, map the files the plan will touch. Apply, commit.
- 4. **Plan.** superpowers:writing-plans, commit; **Fable review the plan** (it also does SDD's pre-flight scan); apply, commit.
- 5. **Execute.** superpowers:subagent-driven-development. Always pass `model`: sonnet for implementers and task reviewers, Fable only for gates. Two review rounds per task at most, then decide and log.
- 6. **Verify, then Fable review the code** from a review package of `<base>..HEAD`. One fixer for all findings. Re-review once, the fix diff only, only if a Critical was fixed.
- 7. **Finish.** Full verification (suite, typecheck, project smoke and golden checks, screenshots of visual changes) once per HEAD; a run recorded in `Verified` for this HEAD is the evidence. Merge only if the order said so: merge master in, re-verify, merge into master locally, remove the worktree per the project's memories. Otherwise leave the branch. Never push or deploy unless ordered.
- 8. **Report** what was built, branch and commit, test count, findings applied and rejected, the owner's open gates. Mark the memory `AUTOLOOP DONE`.
- Findings: apply those you agree with; reject the rest in one written line each.
- ## Stop only for
- - Destructive or scope-changing calls (deleting user work, licences, abandoning the spec): stop and report.
- - Owner-only blockers (credentials, payment, two-human playtest, push): finish everything else, then report.
README.md · README · 34 lines
- # autoloop
- A Claude Code skill for the "I'm leaving this running" loop: write the spec, take every recommendation, Fable review the spec, write the plan, Fable review the plan, execute it, Fable review the code, and finish, with no check-ins until the code is complete, tested and reviewed.
- It builds on the [superpowers](https://github.com/obra/superpowers) skills (brainstorming, writing-plans, subagent-driven-development, using-git-worktrees) and pre-answers every approval gate in them so the run never stalls.
- ## Features
- - Size tiers (S / M / L) so small changes take a short path
- - State in a project memory (`AUTOLOOP RUNNING`, `Done` / `Next`) so a long run resumes after context compaction
- - A bounded Fable review prompt (`review-prompt.md`): at most 10 findings, concrete scenario plus exact fix
- - Sonnet for implementers and per-task reviews, Fable only for the gates
- - Never pushes or deploys unless the order says so
- ## Requirements
- - Claude Code with the superpowers plugin
- - Access to the Fable model (or edit the skill to name another reviewer model)
- ## Install
- ```bash
- git clone https://github.com/GregProuty/autoloop-skill ~/.claude/skills/autoloop
- ```
- ## Use
- ```
- /autoloop build the extinguisher spike, merge when done
- ```
- or say you're leaving it running and want the full spec-plan-build loop without check-ins.
- MIT license; see [LICENSE](LICENSE).
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.