autopilot
Skill0 starsChanged 4 months agoMITView on GitHub
Runs multiple Sprints autonomously as developer and product owner. Executes plan-run-verify-done cycles guided by VISION.json and DESIGN_PRINCIPLES.json, stopping at milestones for user review.
81 other projects use a version of this.
Adopted by
Most recently taken up by
- GongYuanCaiJi/jobsalvotheir own version, first seen Nov 16, 2026
- BoredHF/albertskillstheir own version, first seen Nov 16, 2026
- Yomudogly/oh-my-agytheir own version, first seen Nov 16, 2026
- noelceta/content-enginetheir own version, first seen Nov 16, 2026
- kele-dev/Claude-Orchestrator-Skilltheir own version, first seen Nov 16, 2026
- mkritter3/codex-paired-superpowerstheir own version, first seen Nov 15, 2026
Used with
Projects with it also use review-bugbot, rename-chat and 10 more.
Other files people use alongside this one.
Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
0mlkrizzz655335vSkillreview-bugbot
缺陷导向代码审查。用 Bugbot 子代理找 bug。
Also in 3 projects that use this
0mlkrizzz655335vSkillrename-chat
重命名当前聊天。按聊天内容改标题。"disable-model-invocation: true
Also in 3 projects that use this
Tsumugii24Skillsplit-to-prs
Split current work into small reviewable PRs. Use when the user asks to split a chat, set of changes, branch, or PR.
Also in 4 projects that use this
PandelisZSkillupdate-cli-config
View and modify Cursor CLI configuration settings in ~/.cursor/cli-config.json. Use when the user wants to change CLI settings, configure permissions, switch approval mode, enable vim mode, toggle display options, configure sandbox, or manage any CLI preferences.
Also in 3 projects that use this
32103210Skillmigrate-to-skills
Convert 'Applied intelligently' Cursor rules (.cursor/rules/*.mdc) and slash commands (.cursor/commands/*.md) to Agent Skills format (.cursor/skills/). Use when you want to migrate rules or commands to skills, convert .mdc rules to SKILL.md format, or consolidate commands into the skills directory.
Also in 3 projects that use this
32103210Skillupdate-cursor-settings
Modify Cursor/VSCode user settings in settings.json. Use when you want to change editor settings, preferences, configuration, themes, font size, tab size, format on save, auto save, keybindings, or any settings.json values.
Also in 3 projects that use this
- +1 more from 32103210
- +1 more from 0mlkrizzz655335v
Other versions81 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 81
- Yeachan-Heo/oh-my-claudecode★ 41ksame name
- Yeachan-Heo/oh-my-claude-sisyphus★ 41ksame name
- Yeachan-Heo/oh-my-codex★ 34ksame name
- SethGammon/Citadel★ 927same name
- H-mmer/pentest-agents★ 815same name
- Harry-kp/vortix★ 757same name
- nick-vels/skills★ 599same name
- yan-labs/yan-skills★ 307same name
- kentcdodds/kcd-skills★ 64same name
- Kanevry/session-orchestrator★ 53same name
71 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository3 other files it gives its agents.
- designSkill
- project-initSkill
- sprintSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
13 rules. 0 are common, 3 are distinctive, and 10 are very specific to this project.
0
Common
Most of the 72 edited versions we read keep these.
3
Distinctive
This author's own choices. Worth a look.
10
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 21
If VISION.json or DESIGN_PRINCIPLES.json do not exist, do not proceed.
- Distinctive: line 38
When a command is invoked, read the corresponding reference file before taking any action.
- Very specific: line 45
Dependency boundary: read the dependencies object in ROADMAP.json.
- Distinctive: line 47
End of roadmap: all Sprints complete
- Very specific: line 55
Read docs/ROADMAP.json for overall progress (progress.percentage, done / in_progress / remaining, current Sprint, next milestone).
- Very specific: line 56
Read the most recent docs/sprint-logs/*/decisions.json files for key decisions.
- Very specific: line 57
Read the most recent milestone's docs/sprint-logs/*/compromises.json (if present) and summarize compromises by severity (high first).
- Very specific: line 59
If docs/DESIGN/ exists, show ADR counts by status (accepted / tentative / superseded) and the number of open questions (tentative: true …
- Distinctive: line 71
Never skip milestones: Always stop at milestone boundaries.
- Very specific: line 108
Idempotency requirement: every invocation re-reads docs/ROADMAP.json, the milestone's compromises.json, and docs/DESIGN/ from scratch.
- Very specific: line 128
When a routed ②/③ item lands in the backlog rather than directly in a Sprint, tag it with its kind (② → bug or enhancement, ③ → enhancement …
- Very specific: line 170
Run design refresh internally → check ADR contradictions and broken links.
- Very specific: line 183
Offers default to No: If autopilot offers to tidy an existing ROADMAP.json to the new schema, the default is to leave it alone.
“Common” is measured: kept by at least 6 in 10 of the 72 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Autopilot
Runs multiple Sprints autonomously, acting as both developer and product owner. Guided by docs/VISION.json, docs/DESIGN_PRINCIPLES.json, and (when present) docs/DESIGN/. Stops only at milestone boundaries for user review.
Prerequisites
| File | Required | Purpose |
docs/ROADMAP.json | Yes | Sprint definitions and execution order |
docs/VISION.json | Yes | Product purpose, target users, success criteria |
docs/DESIGN_PRINCIPLES.json | Yes | Decision-making rules for autonomous execution |
docs/DESIGN/ | Optional | Load-bearing architectural decisions (managed by design skill). When present, ADRs are treated as binding constraints on autonomous decisions. |
If VISION.json or DESIGN_PRINCIPLES.json do not exist, do not proceed. Suggest one of:
- Complex / non-trivial systems (recommended): invoke the
designskill viadesign start. It produces VISION, DESIGN_PRINCIPLES, ANDdocs/DESIGN/through guided dialogue. - Simple projects: help the user create VISION and PRINCIPLES inline using
references/vision-template.mdandreferences/principles-template.md.
For complex systems with many cross-cutting decisions, the inline templates are not enough — use design instead.
Commands
| Command | Description | Detail |
autopilot setup | Create VISION, DESIGN_PRINCIPLES, and required artifacts | See references/autopilot-setup.md |
autopilot start | Begin autonomous multi-sprint execution | See references/autopilot-start.md |
autopilot review | Triage user requests after a milestone and route them to the right mechanism (idempotent) | See ## Review Mode below |
autopilot status | Show progress and recent decisions | See below |
autopilot help | Show command list and usage guide | See references/autopilot-help.md |
When a command is invoked, read the corresponding reference file before taking any action.
Milestone Detection
A milestone boundary is any of the following (checked in order, use the earliest match):
- Explicit milestone marker in ROADMAP.json: a Sprint with
"milestone": true - Dependency boundary: read the
dependenciesobject in ROADMAP.json. If the next Sprint has no dependency on the current Sprint (a new independent track begins), treat the current Sprint as a milestone. Determined by explicit dependency declarations, not by inference. - Every 3 Sprints as a fallback if no explicit milestones or dependency boundaries are found — prevents unbounded execution
- End of roadmap: all Sprints complete
These same rules define the detail horizon used by sprint roadmap (rolling-wave planning): sprints up to the first milestone boundary are generated in full detail, sprints beyond it are coarse placeholders elaborated just-in-time at each milestone batch (see ../sprint/references/sprint-roadmap.md → "Detail horizon" and references/autopilot-start.md → "Batch elaboration"). A coarse sprint (detail_level:"coarse" / empty stories) cannot be run or marked done until elaborated.
autopilot status
Show the state of the most recent autopilot run. Read-only — no active session required.
- Read
docs/ROADMAP.jsonfor overall progress (progress.percentage,done/in_progress/remaining, current Sprint, next milestone). - Read the most recent
docs/sprint-logs/*/decisions.jsonfiles for key decisions. - Read the most recent milestone's
docs/sprint-logs/*/compromises.json(if present) and summarize compromises by severity (highfirst). If absent, show "妥協なし / 記録なし". 3.5. Read the recent milestones'skill-retrospective.md(if present). Surface any Deferred SKILL-diff proposal that has been carried for 2+ milestones — that is the self-audit item 1 基準 for a stalled improvement loop, and it is reported to the PO here. If the most recent milestone has no retrospective at all, note "skill-retrospective 未生成". - If
docs/DESIGN/exists, show ADR counts by status (accepted / tentative / superseded) and the number of open questions (tentative: trueitems), perdesign status. - Inspect
.claude/autopilot-*.lockfiles andgit worktree list | grep autopilot/perreferences/autopilot-operations.md. - Present a one-screen view: roadmap progress + current Sprint, next milestone, last completed Sprint, recent compromises (by severity), any stalled SKILL-diff proposals (retrospective Deferred carried 2+ milestones), DESIGN/ ADR summary (if present), active sessions, remaining worktrees + merge status, key decisions, and any drift warnings or failure logs.
Backward compatibility: on a project that predates these artifacts, missing fields render as "-" / "N/A" rather than erroring (see "Backward compatibility" below).
Important Behaviors
- VISION, PRINCIPLES, and DESIGN/ are the authority: Every autonomous decision must be justifiable by referencing one of these documents. ADRs in
docs/DESIGN/adr/are binding constraints — autonomous decisions that contradict an accepted ADR must escalate to the user, not proceed. If none address the question, default to the simplest approach and log why. - 8-Guard Done Judgment: Before marking any Story as
done, autopilot must apply all 8 guards defined in../sprint/references/sprint-done-judgment.md(the canonical definition;references/autopilot-done-judgment.mdis a pointer to it). The mechanical guards run once by machine (hooks/run-guards.py→guards-run.json); the model applies policy on the hits. Any failed guard moves the Story toneeds_user_review, notdone. Guards are evaluated atsprint verifyand machine-drift-checked atsprint done; autopilot itself only reads the recorded results (never re-derives them). - priority_rule 9 exception scope is strict: The exception clause (障害シナリオへの限定) requires explicit障害シナリオ identifiers (
kill-9/停電/Shamir-unseal/ネットワーク遮断/disk-full/OOM/プロセスクラッシュ) in thereview_reason. autopilot rejects exception claims without these markers and falls back to the normal real-VM smoke requirement. - Mock-mode does not satisfy real-VM smoke: Tests that use
MOCK=true,--fake-*flags,DRY_RUN=1, in-process FakeCore / InMemoryStore, or*_fake_*: trueAnsible defaults do not count toward priority_rule 9 "dev VM 実機 smoke" requirement. A separate real-mode smoke is required. - Never skip milestones: Always stop at milestone boundaries. The user's review is the alignment mechanism.
- Forbidden actions are split into immediate-stop and notify-after: Autopilot must obey
## Constraints and Forbidden Actionsbelow. Immediate-stop categories (AC tampering, destructive git, ADR violation, falsedone) halt the run; notify-after categories (test weakening, error swallowing, type-safety relaxation) are recorded tocompromises.jsonand surfaced at the milestone. - Milestone artifacts are mandatory: On reaching a milestone, autopilot writes
docs/sprint-logs/{SprintID}/compromises.json(perreferences/COMPROMISES_SCHEMA.json), thencomprehension-report.md(perreferences/comprehension-report-template.md), thenskill-retrospective.md(perreferences/skill-retrospective.md), and tells the user to read the comprehension report beforeautopilot review. - Failure → SKILL diff, not just a task fix: The
skill-retrospective.mdwritten at each milestone forces item 1 of the self-audit (docs/skills-self-audit.md): for every compromise, overlooked verifier finding, and review re-open in the batch, classify whether the root cause is task-local or a defect in a SKILL/harness/hook, and for a defect propose a diff or record an explicit reason for deferral (never leave it blank). Advisory — it never blocks the milestone, and autopilot never edits a SKILL from here; the proposal goes to the PO through the self-audit flow (SKILL revision is load-bearing). - Independent verifier under
--auto: When autopilot drivessprint verify(always--auto), it auto-enables--with-verifier— a separate read-only Claude session (../sprint/references/verifier-agent.md) re-checks AC, forbidden categories, ADR conformance, and compromise completeness. Itsverification-report.jsonis the trust source: where it disagrees with autopilot's self-report, the verifier wins and the gap is recorded asoverlooked_by_autopilot. - Drift logging, not drift blocking: Log decisions that seem to conflict with VISION/PRINCIPLES; surface them at milestone review. Doc-staleness and VISION-drift health checks (operations.md) are advisory at milestone, never blocking.
- Preserve user agency: The user can always interrupt autopilot. Pause and respond before continuing.
- Context management: The main autopilot conversation is the orchestrator and the sole spawner — Story-level parallelism requires fan-out from the top level, and a sub-agent cannot spawn sub-agents, so a whole Sprint is never wrapped in one sub-agent (that would serialize its Stories). Sprints are sequential by nature (each builds on the prior increment), so no Sprint-level parallelism is lost. Context stays thin by delegation, not by wrapping: every heavy phase (plan/elaboration, Story implementation, review-fix, verify) runs in a leaf sub-agent that returns a structured summary, while the main conversation retains only wave structure, branch names, decision summaries, and status. Per-sprint working detail lives in leaf sub-agents and on disk (
docs/sprint-logs/,docs/ROADMAP.json), so it does not accumulate across the batch. - Incremental commits + branch lifecycle: Each Sprint on
autopilot/{base-branch}/{SprintID}; merged branches are deleted (local + remote). Only unmerged branches survive. Per-branch locking lets multiple branches run concurrently. Full procedure:references/autopilot-operations.md.
Constraints and Forbidden Actions
Autopilot runs unattended, so it must never trade away correctness for a green run. Forbidden actions fall into two tiers by how much damage continuing would do:
| Category | Examples | On detection |
| Effectively disabling a test | adding it.skip / xtest / @pytest.mark.skip / t.Skip; rewriting to expect(true).toBe(true) | Notify after completion |
| Weakening an assertion | toEqual → toBeTruthy; concrete value → any value; removing assertions | Notify after completion |
| Swallowing errors | newly added try { ... } catch {}, // @ts-ignore, # noqa | Notify after completion |
| Abandoning type safety | changing to any, abusive as casts | Notify after completion |
| Deleting / loosening acceptance criteria | removing an AC from ROADMAP.json, softening its wording | Immediate stop |
| Destructive git | push --force, reset --hard origin, branch deletion of unmerged work | Immediate stop |
| False status in ROADMAP / verification-results | writing status: "done"/"pass" while tests fail (incl. rationalizing a real failure as "pre-existing"/"out of scope") | Immediate stop — deterministically blocked by the machine verdict gate (../sprint/references/verify-execution.md) |
| Implicit ADR violation | implementing against an accepted ADR's Decision and justifying it only via decisions.json | Immediate stop (require an ADR amendment) |
Decision rule:
- Immediate stop = the action breaks the premise of all later work (AC tampering, git destruction, ADR violation, false done). Continuing is meaningless — halt and escalate to the user now.
- Notify after completion = a local concession that stays reviewable later. Record it to
docs/sprint-logs/{SprintID}/compromises.json(perreferences/COMPROMISES_SCHEMA.json) and keep going; stopping the whole run for a local concession costs more than it saves.
The independent verifier (--auto always enables it) re-scans the diff for these same categories; an item it finds that autopilot missed is recorded with overlooked_by_autopilot: true.
Three defense layers enforce this table: L1 this prompt (autopilot self-restraint), L2 the sprint verify forbidden-degradation diff scan (../sprint/references/test-discipline.md Rule 6 — pattern-run by machine via hooks/run-guards.py → guards-run.json, with the model applying policy on the hits), and L3 the optional PostToolUse hooks hooks/forbidden-action-guard.py (blocks test-disabling edits during an autopilot run) and hooks/verification-integrity-guard.py (blocks recording a pass over a machine-recorded failure). L3 ships with the plugin; see hooks/README.md. The "false status" row above is additionally enforced deterministically: under --auto, sprint verify derives test status from the machine verdict (hooks/run-verify.py → verify-run.json, real exit codes), and sprint done refuses to complete a Sprint whose overall_machine_status is not pass — the model cannot author a passing result over a failing run (../sprint/references/verify-execution.md, test-discipline Rule 9).
Review Mode
autopilot review is the idempotent triage command invoked after a milestone. It does not implement fixes itself — it classifies each user request and routes it to an existing mechanism. The user only has to remember autopilot review; Claude does the classification.
Idempotency requirement: every invocation re-reads docs/ROADMAP.json, the milestone's compromises.json, and docs/DESIGN/ from scratch. It holds NO internal state between calls. This lets the user loop touch the app → notice something → autopilot review → fix → touch again any number of times at one milestone boundary, and across milestones, with identical behavior each time.
Flow
- Comprehension gate — confirm
docs/sprint-logs/{SprintID}/comprehension-report.mdexists for the milestone's Sprints. If missing, generate it (perreferences/comprehension-report-template.md) before continuing, and ask the user to read it first. - Load context — ROADMAP.json + the milestone's
compromises.json+ recent Sprint status +docs/DESIGN/(if present). - Collect requests — take all of the user's requests at once (batch).
- Classify each request into ①–④ using the decision tree below; for ④ with DESIGN/, also surface the ADR impact.
- Present the classification table to the user for approval.
- Route each approved request to its mechanism; update ROADMAP.json (and ADRs if needed); leave the project ready for the next
autopilot start.
The four classes and where each is routed
| # | Class | Test | Routed to |
| ① | AC violation — an existing AC was not actually met | sprint verify failure handling → re-open the Sprint (see below) → sprint run to fix | |
| ② | Out-of-AC small fix — AC met, minor UX/detail, 1–2 tasks | sprint fix (a fix-Story; placement chosen by the user, see below) | |
| ③ | Out-of-AC new scope — Story-sized or larger new work | sprint idea → picked up by the next sprint plan | |
| ④ | Direction change — ROADMAP-wide / load-bearing | DESIGN/ present: design refresh → design adr → sprint roadmap. DESIGN/ absent: sprint roadmap re-run |
When a routed ②/③ item lands in the backlog rather than directly in a Sprint, tag it with its kind (② → bug or enhancement, ③ → enhancement or feature; see ../sprint/references/ROADMAP_SCHEMA.json → backlog_kind) so later elaboration keeps this classification instead of re-deriving it.
Decision tree (apply per request)
Q0: Does DESIGN/ have a relevant ADR or VISION item? YES → note ADR impact alongside the class. NO → continue.
Q1: Should an existing AC already have covered this? YES → ① AC violation. NO → Q2.
Q2: Does it complete in 1–2 small tasks? YES → ② small fix. NO → Q3.
Q3: Does it fit the existing roadmap direction + ADRs? YES → ③ new scope. NO → ④ direction change.Conservative bias when a boundary is unclear — always lean toward the costlier-to-miss side:
- ① vs ② → lean ① (missing an AC violation is worse than an extra check).
- ② vs ③ → lean ③ (don't pollute a done Sprint with creep).
- ③ vs ④ → lean ④ (don't miss a load-bearing decision).
① Sprint re-open detail
| Change | Action |
| Sprint status | done → in_progress |
| Affected AC status | pass → fail, add reopened_at: <timestamp> |
| Fix work | append fix tasks to the Story, or add a fix-Story |
| Log | write docs/sprint-logs/{SprintID}/reopen.json (per ../sprint/references/REOPEN_SCHEMA.json), triggered_by: "milestone_review" |
| Progress | recompute progress.done |
(Re-opening a past Sprint by explicit user request uses the same flow but triggered_by: "manual_retroactive". Autopilot never re-opens a past Sprint on its own — see ## Backward compatibility.)
② Fix-Story placement detail
Adding a Story to a done Sprint is normal in agile. Let the user choose placement:
| Placement | Behavior |
| Most recent done Sprint (default) | add to that Sprint's stories with added_in_review: "<milestone>" |
| Next Sprint | add to the next (or a new) Sprint |
The added_in_review field keeps "added after the fact" distinguishable from "originally planned" in history.
④ Direction-change detail (DESIGN/ present)
autopilot reviewclassifies ④.- Run
design refreshinternally → check ADR contradictions and broken links. - Present the impact to the user.
- On approval,
design adradds a new ADR or supersedes an existing one. - Re-run
sprint roadmap(if needed); update downstream Sprints.
Without DESIGN/, ④ is just a sprint roadmap re-run. Full handoff contract: design/SKILL.md → "autopilot review ④ direction-change handoff".
Backward compatibility
This skill must run cleanly on projects created before these mechanisms existed:
- Never auto-convert existing files. ROADMAP.json / VISION.json / DESIGN/ / sprint-logs/ are read as-is. No
migratecommand. - New fields are optional. A Story without
added_in_review/reopened_atis read as "originally planned" / a normalpassAC. Missingcompromises.json⇒ "妥協なし"; missingcomprehension-report.md⇒ generated on demand atautopilot review, never back-filled for past Sprints; missingverification-report.json⇒ "verifier 未実行"; missingskill-retrospective.md⇒ "skill-retrospective 未生成", generated on demand at the next self-audit, never back-filled for past milestones. - No retroactive rewrites. Autopilot does not re-open or re-grade a past
doneSprint on its own. Only an explicit user instruction does, recorded astriggered_by: "manual_retroactive". - Offers default to No. If autopilot offers to tidy an existing ROADMAP.json to the new schema, the default is to leave it alone; only an explicit yes proceeds.
Reference Files
references/autopilot-setup.md—autopilot setupcommand flow (new project / existing project, DESIGN/ detection, Backfill handoff)references/autopilot-start.md—autopilot startcommand flow (pre-flight, prototype review, sprint loop, milestone demo, cleanup)../sprint/references/sprint-done-judgment.md— canonical 8-guard done judgment (Guard 1–8) applied before any Story can be markeddone(references/autopilot-done-judgment.mdis a pointer to it)references/autopilot-operations.md— branch locking, worktree cleanup, sprint branch deletion, milestone health checks (doc staleness, VISION drift)references/COMPROMISES_SCHEMA.json—compromises.jsonschema (notify-after compromises recorded at a milestone)references/comprehension-report-template.md—comprehension-report.mdtemplate + writing guide (generated at every milestone)references/skill-retrospective.md—skill-retrospective.mdtemplate + guide: failure/rework → SKILL diff proposal (self-audit item 1), generated at every milestone../sprint/references/verifier-agent.md— the independent read-only verifier sub-agent spec (auto-enabled under--auto)../sprint/references/VERIFICATION_REPORT_SCHEMA.json—verification-report.jsonschema (verifier output; trust source for compromises)../sprint/references/REOPEN_SCHEMA.json—reopen.jsonschema (① Sprint re-open in Review Mode)references/getting-started.md— New project setup guide (with specs / without specs / existing project)../design/references/VISION_SCHEMA.json— VISION.json schema and example. Owned by thedesignskill (the authority that generates VISION). autopilot setup's inline VISION generation reads it but a stricter version is produced bydesign.../design/references/DESIGN_PRINCIPLES_SCHEMA.json— DESIGN_PRINCIPLES.json schema and example. Owned by thedesignskill for the same reason.references/vision-template.md— VISION setup guidelines (question prompts)references/principles-template.md— DESIGN_PRINCIPLES setup guidelines (question prompts)references/autopilot-help.md— help (command list and usage guide)
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

