autopilot
Drives an idea from any pipeline entry point — a cold raw prompt included — to one open MVP pull request in a single autonomous run, pausing only at a single…
Skill by specscore
Chalk Labs
Need help deploying autopilot?
Chalk Labs sets up agents like this inside real teams: connected to your tools, tested on your data, with a human in the loop where it matters.
Install
Works with Claude Code, Claude apps, Codex, Cursor, Gemini CLI, GitHub Copilot, VS Code, OpenCode, Goose, Windsurf. Read the repository before you give any tool access to your data.
npx agentmarket install skill/specscore-autopilot
Asks which of your tools to add it to (Claude Code, Codex, Cursor, VS Code, Windsurf, Gemini CLI, Claude Desktop) and shows the change first. See https://agentmart.dev/cli.
Get the weekly diff
The weekly diff: one email a week with new agent tools, notable updates and grade changes, plus security alerts for tools you follow. Double opt-in, one-click unsubscribe.
About
Use cases: devops, data, agent-building
Reviews
Write a reviewNo reviews yet. Installed it? Reviews come from verified installers only. How reviews work
Permissions
What this skill can do and what it asks for, from its SKILL.md. Checked Oct 25, 2026.
- Tools
- No tools found
- Secrets
- None asked for
- Install scripts
- None found
Network hosts it talks to 1
codeload.github.com· found in install command
Files it touches
~/.claude/skills/autopilot
What's inside
- Skill name
autopilot- Description
- Drives an idea from any pipeline entry point — a cold raw prompt included — to one open MVP pull request in a single autonomous run, pausing only at a single Idea checkpoint and halting on genuine anomalies. Thin orchestrator: it detects the furthest-along artifact, then chains the existing producer skills ideate → specify → plan → implement → pull-request, arming a run-scoped autonomy signal that releases the human-approval gates and switches the producers into decide-and-record. It re-implements no producer logic and reuses the approval-autonomy implement/plan gate layer unchanged. Publish ceiling is local commits + one PR — never merge, deploy, or ship. Trigger: "do it autonomously", "/autopilot", "/autopilot <slug>", or "autonomously" appended to any ideate/specify/plan/implement request.
- License
- MIT
- Instructions
- 128 lines in SKILL.md · view on GitHub
Files (1)
- SKILL.md
SKILL.md (first lines)
# Autopilot Turn a single trigger into an end-to-end autonomous run: from a raw prompt (or any partial artifact) to one open MVP pull request, with no per-question stops and exactly one human checkpoint at the crystallized Idea. Autopilot is a **thin orchestrator** — it drives the existing `ideate → specify → plan → implement → pull-request` skills, arms the run-scoped autonomy signal defined in [autonomy-autopilot.md](../shared/autonomy-autopilot.md), and hands back the trail. It owns orchestration, the Idea checkpoint, the publish ceiling, and the handback — nothing else. ## Hard Gate <HARD-GATE> Autopilot MUST NOT: - Re-implement any producer skill's logic, gates, or artifact writes — it only invokes them. - Mask any gate other than a `type: human` approval entry. `type: ai` / `type: deterministic` reviewers, `specscore spec lint`, and `implement`'s conflict detection ALWAYS run and can halt the run. - Skip the `confirm_idea` checkpoint when the run auto-creates an Idea and `autonomy.autopilot.confirm_idea` is not `false`. - Publish past its ceiling: local commits + exactly ONE pull request. It MUST NOT merge, deploy, invoke `ship`, open a second PR, or retry a failed PR delegate. - Auto-invoke `specstudio:verify` or `specstudio:ship`. If a stage's quality gate returns `Issues Found`, or an anomaly halts `implement`, autopilot STOPS and hands back the blocker — it never releases a quality gate to keep moving. </HARD-GATE> ## When to Use - The user says "do it autonomously", `/autopilot`, `/autopilot <slug>`, or appends "autonomously" to any pipeline request — **even on the very first prompt**, before any artifact exists. - The user wants to reach an MVP without answering per-stage clarifying questions or approving each gate, accepting one Idea checkpoint and a decision log they review afterward. **Refuse only** when the input is a genuinely empty ask — no prompt and no resolvable artifact to act on. ## Entry-Point Detection Autopilot has **no fixed precondition** — it starts wherever the work currently is. Resolve the furthest-along existing artifact for the work, then run every remaining stage: | Input state | Autopilot runs | |---|---| | Raw prompt, no artifact | ideate → specify → plan → implement → pull-request | | `Draft` / `In Review` Idea | ideate (finish) → specify → plan → implement → pull-request | | `Approved` Idea | specify → plan → implement → pull-request | | `Approved` Feature, no Plan | plan → implement → pull-request *(or implement directly, per `implement`'s Feature-sourced mode)* | | `Approved` Plan | implement → pull-request | Resolution steps: 1. If the trigger names a slug (`/autopilot <slug>`), resolve that artifact; otherwise infer the work from the prompt and scan `spec/ideas/`, `spec/features/`, `spec/plans/` for a matching in-flight artifact. 2. Pick the furthest-along artifact in the pipeline order `idea < feature < plan`. Enter at the stage that produces the **next** artifact. 3. The trigger phrase **is** the consent to auto-create and auto-approve every artifact the run produces (subject only to the `confirm_idea` checkpoint). There is no separate "must be an approved Idea first" gate. 4. If nothing resolves and there is no prompt to ideate from, refuse and say what's missing. ## The Run Autopilot performs the run as an ordered drive over the producer skills. It does **not** reimplement any of them; each stage is the existing skill invoked in the current session. 1. **Arm.** Establish the run-scoped autonomy signal ([autonomy-autopilot.md](../shared/autonomy-autopilot.md)) at run scope — **once** for the whole run (see Single Arming). Resolve `autonomy.autopilot` (`publish_ceiling`, `confirm_idea`, `stop_on`) across the scope ladder. 2. **Disclose.** Emit the single up-front disclosure message (see the Disclosure section) before any stage runs. 3. **Drive the stages** from the resolved entry point, in order, advancing to the next stage only after the current stage completes successfully: - **ideate** (if entry ≤ Idea) → then honor the `confirm_idea` checkpoint (see the Idea Checkpoint section). - **specify** (if entry ≤ Feature) — its `gates.feature.approved` `type: human` entry is masked by the run-scoped signal; the `type: ai` reviewer still runs. - **plan** (if entry ≤ Plan) — its human review gate is masked; the baseline reviewer and P-001 coverage still hold. - **implement** — `implementation.pre_commit` is autonomous per `approval-autonomy`; `implementation.pre_push` is masked here (the push happens at the PR stage). Anomaly-halts still stop the run. - **pull-request** — open exactly one PR (see the Publish Ceiling section). 4. **Hand back** the trail (see the Handback section). At every stage, a masked `type: human` entry is released by the reviewer-gates runner's Step 1.6; the producers take their decide-and-record branch instead of asking clarifying questions. Autopilot itself asks nothing except the one `confirm_idea` checkpoint. ## Single Arming The trigger arms the run-scoped autonomy signal **once**, for the remainder of the run. Autopilot does NOT re-arm at each stage boundary — crossing from `specify` into `plan` into `implement` requires no fresh signal. The one exception is inherited, not owned: if `implement` hits an **anomaly-halt** (sibling conflict, BLOCKED subagent, unresolved lint, source-Feature drift), that is a genuine stop under `approval-autonomy`'s rules, and `implement` requires the explicit `continue` re-arm before resuming. That anomaly re-arm is `implement`'s, scoped to its own stage — it is not a per-stage gate autopilot introduces. A subsequent autopilot run starts unarmed and re-reads config from scratch. ## Anomaly Halts Autonomy masks *human approval*, never *quality*. Autopilot stops and hands back the blocker whenever: - a stage's `type: ai` / `type: deterministic` reviewer returns `Issues Found`; - `specscore spec lint` fails and `--fix` does not resolve it in one pass; - `implement` reports any anomaly in its halt set (per `approval-autonomy`); - a `stop_on` decision class fires (`conflict` is always in the set and not removable). On any halt, autopilot names the specific cause, performs no auto-resolution, and does not advance. The user fixes it and re-triggers (or issues `implement`'s `continue` re-arm for an implement-stage anomaly). ## Idea Checkpoint (`confirm_idea`) When the run **auto-creates an Idea** (entry at or before the Idea stage), autopilot pauses **exactly once** — the single human-approval stop of an otherwise unbroken run. This is the cheapest, highest-leverage place to bound drift: a raw one-line prompt is the most ambiguous possible input, and a 30-second Idea read catches divergence before it propagates downstream. - **Default on.** With `autonomy.autopilot.confirm_idea` unset or `true`, after `ideate` crystallizes the Idea autopilot presents it and waits for explicit approval before invoking `specify`. The `idea.approved` human review gate is **not** masked at this checkpoint — it is deliberately preserved. Recognize approval with the standard explicit-approval phrase set; a vague positive gets one confirmation question. - **Opt-out.** With `confirm_idea: false`, the checkpoint is masked like every other human gate and the cold-start run is fully unbroken. - **Not applicable when the Idea pre-exists.** When the run enters at or after an **already-`Approved` Idea**, no checkpoint fires — the human already approved that Idea. The checkpoint is specifically for Ideas the run itself shaped. On explicit approval (or when `confirm_idea: false`), continue to `specify`. On a change request, fold it into the Idea and re-present — the run does not proceed past an unapproved run-created Idea. ## Publish Ceiling Autopilot's ceiling is **local commits + exactly one pull request**, resolved from `autonomy.autopilot.publish_ceiling` (default `pr`): - Commits land during `implement` via the existing `implementation.pre_commit` gate (autonomous per `approval-autonomy`). `implementation.pre_push` is masked in the implement stage — the push happens once, at the PR stage. - After `implement` completes, autopilot invokes `specstudio:pull-request` **exactly once** to push the feature branch and open a single PR (built-in `git push` + `gh pr create`, or the project's configured `pull_request.delegate`). Its `pull_request.pre_dispatch` `type: human` entry is masked; a `type: deterministic` verify gate there is **not** masked and can still block. - `publish_ceiling: commit` stops after local commits (no PR); `publish_ceiling: stage` stops after staging. Autopilot MUST NOT merge, deploy, invoke `ship`, open more than one PR, or retry a failed PR delegate. The push safety floor (`main`/`master`/`release/*` denied by default) still applies — autonomy never weakens it. ## Disclosure Before any stage runs, autopilot emits **one** disclosure message so the user knows exactly what the run will do unattended. It states: - the **resolved entry point** and the **stages** that will run; - that human-approval gates are **masked for the run**, *except* the single `confirm_idea` checkpoint (named explicitly); - that clarifying decisions will be **auto-made and recorded** to each artifact's decision log; - that the run will **commit locally and open one PR**, and will **not** merge or deploy. The disclosure is informational — it is not a per-stage approval prompt. The run proceeds immediately after it. ## Handback On successful completion, autopilot hands back a single summary — the primary review surface, meant to be read first: - the **resolved entry point** and stages run; - the created **artifact paths** (Idea / Feature / Plan); - the local **commit SHAs**; - the opened **PR URL**; - the **aggregated Autonomous Decisions log** — every decide-and-record entry from every stage (Idea `## Key Assumptions` additions + Feature/Plan `## Autonomous Decisions`), collected in one place. Autopilot does **not** auto-invoke `specstudio:verify` or `specstudio:ship`; it MAY recommend the user run them next.
From specscore/specstudio-skills, MIT.
More from specscore/specstudio-skills
- 0.479score
consilium
specscore
Drains captured sidekick (sideline-idea) seeds and produces deterministic verdicts via a 5-stage pipeline: CLI gather → researcher agent → 9-role parallel…
SkillCFreshness: CClaude Code +9 - 0.479score
specify
specscore
Turns an approved SpecScore Idea (or a clear buildable intent) into a SpecScore Feature with requirements and Given/When/Then acceptance criteria at…
SkillCFreshness: CClaude Code +9 - 0.479score
ideate
specscore
Refines raw ideas into SpecScore Idea artifacts through structured divergent and convergent thinking.
SkillCFreshness: CClaude Code +9 - 0.479score
plan
specscore
Turns an approved SpecScore Feature or Idea into a lint-clean Plan artifact at spec/plans/<slug>.md — an ordered list of tasks where each task references one…
SkillCFreshness: CClaude Code +9 - 0.474score
recap
specscore
Consumes an approved SpecScore Feature plus the latest `specstudio:verify` report at HEAD and produces a per-AC drift report at…
SkillCFreshness: CClaude Code +9 - 0.474score
relocate-idea
specscore
Thin wrapper over the `specscore idea relocate` CLI verb.
SkillCFreshness: CClaude Code +9 - 0.474score
score
specscore
Manually re-invokes the configured reviewer-gates dispatch pipeline against one or more SpecScore artifacts (Ideas, Features, Plans), outside the…
SkillCFreshness: CClaude Code +9 - 0.474score
ship
specscore
Terminal skill of the SpecStudio pipeline.
SkillCFreshness: CClaude Code +9 - 0.474score
sidekick
specscore
Captures a sideline idea as a lint-clean seed at spec/ideas/seeds/<slug>.md without derailing the host task.
SkillCFreshness: CClaude Code +9 - 0.474score
verify
specscore
Consumes an approved SpecScore Feature and produces a machine-checkable per-AC verdict report at spec/features/<feature-slug>/_verify/<sha>.md.
SkillCFreshness: CClaude Code +9 - 0.474score
implement
specscore
Turns an approved SpecScore Plan into focused, AC-traceable source-code changes by dispatching one subagent per task in parallel batches computed from the…
SkillCFreshness: CClaude Code +9 - 0.474score
init
specscore
Bootstraps a SpecScore-managed project in one wizard-driven step.
SkillCFreshness: CClaude Code +9
Signals
- Stars
- 0
- Forks
- 0
- Open issues
- 7
- Last commit (this folder)
- Jul 3, 2026 (3 months ago)
- Last release
- Jun 6, 2026 (4 months ago)
From GitHub (fetched Oct 24, 2026).
Details
- License
- MIT
- Repository
- github.com/specscore/specstudio-skills
- First seen
- Oct 24, 2026
- Last checked
- Oct 24, 2026
- Source
- GitHubgithub-graphql
More from specscore
- 0.479score
plan
specscore
Turns an approved SpecScore Feature or Idea into a lint-clean Plan artifact at spec/plans/<slug>.md — an ordered list of tasks where each task references one…
SkillCFreshness: CClaude Code +9 - 0.479score
ideate
specscore
Refines raw ideas into SpecScore Idea artifacts through structured divergent and convergent thinking.
SkillCFreshness: CClaude Code +9 - 0.479score
consilium
specscore
Drains captured sidekick (sideline-idea) seeds and produces deterministic verdicts via a 5-stage pipeline: CLI gather → researcher agent → 9-role parallel…
SkillCFreshness: CClaude Code +9 - 0.479score
specify
specscore
Turns an approved SpecScore Idea (or a clear buildable intent) into a SpecScore Feature with requirements and Given/When/Then acceptance criteria at…
SkillCFreshness: CClaude Code +9 - 0.474score
init
specscore
Bootstraps a SpecScore-managed project in one wizard-driven step.
SkillCFreshness: CClaude Code +9 - 0.474score
implement
specscore
Turns an approved SpecScore Plan into focused, AC-traceable source-code changes by dispatching one subagent per task in parallel batches computed from the…
SkillCFreshness: CClaude Code +9
Done for you
Rather not set it up yourself?
Chalk Labs scopes, builds and runs agents for your team: the task audit, the eval harness, the integrations and the weekly ops report. Fixed scope, priced before we start.
See how we build AI agents →Read the case studies
Chalk Labs is a two-founder AI and growth studio. We build AI agents, CRMs and MVPs for Web3 and AI companies, and have shipped 50+ product and token launches.
- launches shipped
- 50+
- launches shipped
- media placements
- 500+
- media placements
- founders on every account
- 2
- founders on every account
Press we've placed: Forbes · CoinDesk · Decrypt · Cointelegraph · The Block
Rahil Jain, Co-founder, Chalk Labs. Ex-Goldman Sachs engineer. You talk to the person who builds it.
Get autopilot running in your business
Tell us what you want to automate. A founder replies within one working day with a plan, not a pitch deck.
Cite this page
autopilot skill. Agent Mart. https://agentmart.dev/skills/specscore-autopilot (last updated 2026-10-24)