auto-pilot
Intake-first autonomous project execution for Codex. Use when the user wants a short project request turned into a long-running build workflow with minimal interruptions, or explicitly mentions 오토파일럿/autopilot.
- License 许可
- Apache-2.0
- 最近更新
- 2026-11-14
- Stars
- ★ 1.6k
- Forks
- 491
01 前言 Front Matter
- name
- auto-pilot
- summary
- Ask once, lock spec, save state, keep shipping.
- metadata
- { "priority": 8, "promptSignals": { "allOf": [ [ "ap", "build" ], [ "build", "ap" ], [ "ap", "make" ], [ "make", "ap" ], [ "ap", "create" ], [ "create", "ap" ], [ "ap", "app" ], [ "app", "ap" ], [ "ap", "만들어줘" ], [ "만들어줘", "ap" ] ], "anyOf": [ "오토파일럿", "autopilot", "auto-pilot" ], "phrases": [ "오토파일럿", "autopilot", "auto-pilot", "24시간 계속", "with ap", "using ap" ], "minScore": 7 } }
- retrieval
- { "aliases": [ "오토파일럿", "autopilot mode", "auto-pilot mode", "with ap", "intake first execution", "continuous project execution" ], "intents": [ "start an autopilot project", "run intake and keep building", "continue a project with minimal interruptions" ], "entities": [ "spec", "progress", "blockers", "state" ] }
- description
- Intake-first autonomous project execution for Codex. Use when the user wants a short project request turned into a long-running build workflow with minimal interruptions, or explicitly mentions 오토파일럿/autopilot.
03 SKILL.md 正文
Auto Pilot
Use this skill as the umbrella workflow for long-running project execution. Treat it as the manager role for the whole system.
Goal
Turn a short product request into a resumable execution loop that:
- gathers the minimum required inputs once
- locks a written spec and definition of done
- keeps implementing the next highest-priority work
- escalates only for true human-only blockers
Operating Model
If state does not exist yet
- Route to the intake workflow first.
- Ask only for high-impact decisions.
- Infer safe defaults for the rest.
- After intake, detect whether the selected auth or managed services require upfront integration env values.
- If env values are already present, write:
docs/spec.mddocs/design.mdfor user-facing projectsdocs/progress.mddocs/next.mdautopilot/state.jsonautopilot/blockers.jsonautopilot/secrets-status.json
- If env values are still missing, treat the project as being in the normal
setup-secretsphase and write:docs/next.mdautopilot/state.jsonautopilot/blockers.jsonautopilot/secrets-status.json- request the consolidated env payload
- Start implementation immediately only after the required env values are ready.
If state already exists
- Read
autopilot/state.json,autopilot/blockers.json, andautopilot/secrets-status.jsonif it exists. - If
setupStatusispendingorautopilot/secrets-status.jsonis still pending, treat the project as an existing project waiting on setup continuation. - In that setup continuation path, collect only the missing env payload and do not assume
docs/spec.mdordocs/progress.mdalready exist. - Otherwise read
docs/spec.md,docs/progress.md,docs/next.md, anddocs/design.mdwhen relevant. - Read
executionMode,teamRoles,currentOwner,qualityGates, andlastReviewerVerdictfrom runtime state. - Confirm whether there is an active
human-requiredblocker. - If not blocked, continue with the top unfinished task.
- Save results back to the same files after each loop.
Manager Runtime Contract
The manager owns the real orchestration loop.
Do not treat executionMode, teamRoles, or qualityGates as decorative metadata.
Use them to decide which specialist runs next and when a slice can advance.
Before the first specialist step:
- Read the saved
executionMode. - Check whether Codex native sub-agents are actually available in the current runtime.
- If the saved mode is
team-productorteam-litebut native sub-agents are unavailable, downgrade toserial-fallback. - Immediately update:
autopilot/state.jsondocs/progress.mddocs/next.mdso the recorded execution mode matches the real runtime path.
Intake Rules
Ask for only the minimum complete project contract:
- product summary
- target user
- core features
- non-goals for this version
- stack preference or default stack permission
- architecture preset
- auth requirement
- payment requirement
- admin requirement
- deploy target
- data store
- theme preset
- visual vibe
- design direction refinement
- definition of done
Keep the intake compact, but ask exactly one question at a time.
Treat one-question-at-a-time intake as the default behavior, not an optional preference.
For the exact interaction format, follow ../autopilot-intake/SKILL.md.
Execution Rules
- The manager owns user-facing communication and final state updates.
- Do not send work directly to Builder before Planner has defined the current slice.
- Always choose the smallest shippable slice.
- Prefer working software over broad scaffolding with no validation.
- Run relevant verification after each meaningful change.
- Update progress and next-step files after each loop.
- Continue automatically after intake unless the run is waiting on the upfront integration setup phase or a true
human-requiredblocker is active. - Ask the user only when:
- the upfront integration setup phase needs a consolidated env payload
- external account setup is required and no existing key can be provided yet
- payment or production deployment approval is required
- the product direction is ambiguous in a way that changes the build materially
Treat setup-secrets pending as a normal pre-execution phase, not as a blocker entry by itself.
Team Orchestration
Default orchestration is conditional:
team-product- user-facing product
- or
Scalablearchitecture - or auth / payments / admin are required
team-lite- CLI-only, backend-only, library-only, or simple internal tooling
serial-fallback- use the same role order in one loop when specialist sub-agents are unavailable
Role Activation
Planner: always active at project start and again at milestone transitionsArchitect: active only forScalablearchitecture or cross-cutting structure decisionsBuilder: always activeDesigner: active only for user-facing projectsQA: always active before a slice can be marked complete
Specialist Responsibilities
Planner- break the spec into milestones and current-slice acceptance criteria
- draft
docs/next.mdand the initial quality gates
Architect- define shared vs feature boundaries and integration seams
- prevent structure drift, especially under
Scalable
Builder- implement the next smallest shippable slice
- stay inside the planner/architect contract
Designer- strengthen
docs/design.mdinto the binding UI brief - run one explicit visual review after the first UI pass
- strengthen
QA- validate tests, builds, key flows, regressions, and definition-of-done alignment
- leave a pass/fail verdict before completion
Codex Native Sub-Agent Policy
Use Codex native sub-agents when they are available:
Planner,Architect,Designer,QA:defaultBuilder:worker- repo analysis or structural inspection:
explorer
Keep the manager in control of ordering:
- planner first
- architect and designer only when needed
- builder implementation
- QA review
Designer review is mandatory for user-facing UI slices. QA pass is mandatory before a slice is treated as complete.
Actual Execution Sequence
Use this exact sequence on every slice.
Step 1: Planner
- Always run planner first.
- Use
../autopilot-planner/SKILL.md. - Expected planner output:
- planner checkpoint
- current slice
- acceptance criteria
- risks
- required follow-on roles
- QA checks
- Write planner results into:
docs/progress.mddocs/next.mdautopilot/state.json
- Record the handoff with:
python3 ../../scripts/team_checkpoint.py \
--workspace <target-workspace> \
--role planner \
--summary "<planner summary>" \
--planner-checkpoint "<planner checkpoint>" \
--planner-intent "<planner intent>" \
--builder-target "<builder target>" \
--qa-check "<qa check 1>" \
--qa-check "<qa check 2>" \
--next-owner "<Architect or Builder>" \
--next-task "<next current task>"
Step 2: Architect if needed
Run architect only when:
Scalablearchitecture is active, or- planner identifies cross-cutting structure risk
Use ../autopilot-architect/SKILL.md.
Architect refines boundaries before builder starts and updates the same state files.
Record the result with:
python3 ../../scripts/team_checkpoint.py \
--workspace <target-workspace> \
--role architect \
--summary "<architect summary>" \
--architect-guardrails "<architect guardrails>" \
--next-owner "Builder" \
--next-task "<builder task>"
Step 3: Builder
- Builder only runs after planner, and after architect when architect was required.
- If native sub-agents are available, delegate builder work to a
workersub-agent. - If not, execute the builder role directly inside the manager loop.
- Builder must stay inside the planner and architect contract.
- After a meaningful builder step, record progress with:
python3 ../../scripts/team_checkpoint.py \
--workspace <target-workspace> \
--role builder \
--summary "<builder summary>" \
--builder-progress "<builder progress>" \
--next-owner "<Designer or QA>" \
--next-task "<next current task>"
Step 4: Designer if needed
Run designer only for user-facing slices.
- Use
../autopilot-designer/SKILL.md. - Before the first UI slice, designer strengthens
docs/design.md. - After the first UI pass, designer performs one explicit visual review.
- Manager must record the design verdict into:
docs/design.mddocs/progress.mdautopilot/state.json
- Record the result with:
python3 ../../scripts/team_checkpoint.py \
--workspace <target-workspace> \
--role designer \
--summary "<designer summary>" \
--designer-notes "<designer notes>" \
--reviewer-verdict "<designer verdict>" \
--design-review-note "<design review note>" \
--next-owner "QA" \
--next-task "<qa task>"
Step 5: QA
- Always run QA before a slice is marked complete.
- Use
../autopilot-qa/SKILL.md. - QA must return pass or fail.
- If QA fails:
- do not mark the slice complete
- route rework back to builder
- update
lastReviewerVerdict
- Record the verdict with:
python3 ../../scripts/team_checkpoint.py \
--workspace <target-workspace> \
--role qa \
--summary "<qa summary>" \
--reviewer-verdict "<pass or fail verdict>" \
--qa-check "<qa check 1>" \
--qa-check "<qa check 2>" \
--next-owner "<Manager or Builder>" \
--next-task "<next current task>"
Step 6: Manager update
After the final specialist for the slice:
- update
currentOwner - update
currentTask - update
lastPlannerCheckpointwhen planner completes - update
lastReviewerVerdictwhen designer or QA completes - update
docs/progress.md - update
docs/next.md - append or resolve blockers in
autopilot/blockers.json
Use the same checkpoint script to record blockers when needed:
python3 ../../scripts/team_checkpoint.py \
--workspace <target-workspace> \
--role manager \
--summary "<manager status update>" \
--blocker-classification "human-required" \
--blocker-summary "<minimal blocker summary>" \
--blocker-owner-role "<current owner role>"
Serial Fallback Rule
If native sub-agents are unavailable, run the exact same role order in one manager loop:
- planner
- architect if needed
- builder
- designer if needed
- QA
Do not skip roles just because the runtime is serial.
serial-fallback changes the backend, not the quality bar.
Design Execution Rule
For user-facing products, design synthesis is required before the first UI build.
- Create and follow
docs/design.md - Use the
frontend-ui-uxskill as the default design helper - Prefer Figma references when provided
- Otherwise use curated design inspiration sources rather than broad web search
- Keep the base UI bright: white or light-neutral surfaces, compact readable typography, and restrained chrome
- Choose the navigation shell from the project type: landing pages use top headers, dashboard/admin surfaces use sidebars, and mobile-style products use app-header or bottom-tab shells
- Treat theme and vibe as modulation of accent, contrast, spacing tone, and component shape rather than a full override of the base surface system
- Run one explicit post-build design review after the first UI pass
- Treat
docs/design.mdas the designer-owned binding brief
Treat these as user-facing by default:
- landing pages
- marketing sites
- dashboards
- consumer web or mobile apps
- admin interfaces
Skip dedicated design synthesis only for clearly CLI-only, backend-only, or library-only work.
Blocker Policy
retryable
- dependency install errors
- failing tests
- build issues
- small runtime bugs
Behavior:
- retry automatically
- try one or more alternative fixes
- only escalate after retry budget is exhausted
deferable
- placeholder copy
- icon selection
- optional polish
- non-critical analytics
Behavior:
- keep moving with a reasonable default
- write the deferred decision into
docs/next.mdwhen useful
human-required
- API keys or secrets
- OAuth app creation
- payment account setup
- production deployment approval
- ambiguous product direction that changes scope materially
Behavior:
- stop only at the minimum blocking decision
- ask one focused question
- continue immediately after the answer
- write the blocker with
ownerRoleinautopilot/blockers.json
Each blocker entry must include:
ownerRoleclassificationsummary
04 同仓库更多 Skills
worktable-review-with-annotations
Create, inspect, reply to, update, or resolve situated Worktable annotations. Use when feedback belongs to a Doc section, block, text range, or HTML Doc, or when the user asks to address existing review feedback and coordinate a requested artifact edit.
worktable-find-and-synthesize-context
Find, assess, and synthesize context from Worktable Docs and Records. Use when the user asks what a workspace contains, wants workspace-backed research, needs current facts reconciled across artifacts, or requests a cited synthesis without changing workspace state.
worktable-create-or-update-docs
Create, edit, rename, or maintain narrative Worktable Docs. Use for plans, decisions, briefs, notes, research syntheses, and other durable prose when the user requests a new artifact or a safe update to an existing one.
worktable-create-or-manage-records
Organize repeated structured information as durable Worktable Records. Use for research datasets, inventories, directories, pipelines, trackers, and other collections whose items need identity, fields, filtering, links, validation, or individual updates.
worktable-create-or-manage-interactive-html
Create or improve durable visual and interactive artifacts in Worktable. Use for dashboards, trackers, calculators, mini apps, slide presentations, and other requests that benefit from a polished HTML Doc.
worktable-collaborate-threads
Read, answer, and hand off messages through durable Worktable threads with users or other authenticated participants. Use when the user asks to check addressed messages, reply to a participant, continue a thread, request another agent's help, wait briefly for a response, or watch for comments and messages sent to the agent.