autopilot

Autopilot — autonomous idea→ship orchestrator

autopilot flies the idea-to-ship pipeline end to end, autonomously, in character as a grounded founder-persona — productizing the "paste one kickoff and the pipeline runs itself" pattern. It takes a real, grounded niche and hands back a near-finish-line-AIMED first-draft product + an honest validation ledger.

It is an orchestrator: it composes the existing skills and never reimplements them. It invokes ideate, deep-dive, prompt-pack, and build-loop, carries each one's file output into the next, and adds only the connective tissue — the autonomy contract, the corrected gate, the kill-ledger, and an honest hand-off. If you find yourself pasting a sub-skill's procedure into this run, stop: invoke the skill instead.

It states its limits in its own text (below) because the whole point is an honest first draft, not an oversold "finished product."

When to use this

Strong triggers — invoke without asking:

• "Run autopilot" / "fly the whole pipeline" / "take this idea→ship autonomously"
• "Spin up a grounded founder-persona and build it" / "autonomous first-draft from this concept/niche"

Softer triggers:

• You have a grounded niche (real lived pain or real data) and want a first-draft product + an honest validation ledger without hand-running each phase.

Do NOT use this for:

• Driving each phase yourself — use ideate / deep-dive / prompt-pack / build-loop directly.
• A concept with no real grounding — autopilot will not invent demand to fill the gap (see the firewall in Bounds).
• Market validation — out of scope; that is the human/market handoff.
• A settled code change (prompt-pack) or a soundness audit (deep-dive).

Routing tie-breaker: autopilot automates the manual tier. Reach for it to fly the pipeline end-to-end with eyes open on the bounds — not to replace human taste, finish, or market judgment.

What it produces — and its honest bounds (read this before you trust it)

Load-bearing, stated up front so no one reads an autonomous run as more than it is.

Produces: a grounded CONCEPT_BRIEF, a validated + sequenced build pack, a near-finish-line-AIMED first-draft product (designed + machine-validated as far as machines can check), and an honest ledger of what's checked vs. what only a human/market can verify.

The 3 eyes-open limits it does NOT escape:

1. ~80% craft ceiling + a last-mile tail. "Near-finish-line" is the aim, not a guarantee; a correctness/security/taste tail remains for the human to finish.
2. The grounding firewall. Real data to discover the problem and seed the build = yes. A synthetic persona's reaction to the proposed solution never counts as validation. autopilot injects real signal at the front; it does not manufacture demand.
3. Judgment quality isn't cleanly measurable. Its go / iterate / kill calls are a signal a human weighs, never proof. "The system's judgment matches an expert" is non-gating (it's circular — it needs the market truth the system exists to replace).

Market validation is explicitly the human handoff. "Works + looks good + well-judged" ≠ "people want it." autopilot is a force-multiplier toward a finishable first draft — not a finished-product factory and not a market validator.

The pipeline it flies (compose, never copy)

Hard rule: invoke each skill and carry its file forward; never inline its content.

1. ideate → a grounded CONCEPT_BRIEF.md — forced success-metric + kill-criterion, and (via its conditional-design forcing-function) a substantive design direction when feel is load-bearing.
2. deep-dive → validates the brief's load-bearing claims with honest confidence + a ground-truth tally; its hand-back is folded into the brief (verdict, what's verified vs. still-a-bet).
3. prompt-pack → sequences the validated/gated scope into self-contained build prompts; carries execute-discipline.
4. Gate-impact pass → before each unit's build opens, enumerate the gates, frozen constraints, and conventions that unit's planned surface will trip, and clear them all in one pass. The mechanism lives in prompt-pack; autopilot's job is that the pass happens. Not optional — it is the phase gate between planning and building.
5. build-loop → drives each shippable unit to near-finish-line craft on two co-equal tracks — the objective machine facts AND, when feel is load-bearing, the mandatory multi-pass visual design loop.

Phase-by-phase orchestration (what each invocation gets, what carries forward, where the gates are): references/pipeline-playbook.md — the gate-impact pass sits between its Phase 3 and Phase 4.

The autonomy contract

• Run in character as the grounded persona. Answer every phase gate the founder can answer — the brain-dump, the selection rubric, the divergence ranking, the pressure-test verdict, the convergence locks — in-character, without stopping for a human.
• But never fake a human-only gate. A gate that genuinely needs a human or real-world signal (real use, a taste sign-off, a market/paste-into-prod test) is emitted honestly, not auto-passed in-character. (Execute-discipline, below — the experiment's sharpest failure was a run that narrated a gate it had actually abandoned.)
• Narrate the phases. Say which skill is running and what file it produced, so the run is legible and resumable.

Execute-discipline (inherited from prompt-pack) — never fake a gate

Build only the validated/gated scope — never extra breadth to look complete. If a phase's gate needs a human or real-world signal you don't have, STOP and emit the gate (name what's unverified and who must clear it); never fake it, render a "passed"/inert gate, or build past it. autopilot inherits this from prompt-pack and enforces it across the whole run — the autonomy makes it easy to rationalize past a gate, so this is the rule that keeps an autonomous run honest.

The corrected gate (proceed / park)

• Gate on real-use + checkable craft: does it build, run, pass its flows and (when load-bearing) its design bar via build-loop? Is the output a finish-the-last-mile takeover, not a from-scratch rebuild?
• Non-gating signal: "the system's judgment matches an expert." Report it as a signal, never as the gate (it's circular). An expert backtest is agreement-by-construction.
• Parking is a valid, honest outcome. Say so and stop — then log it (kill-ledger, below).

Design, carried through (per ideate conditional-design + build-loop's visual loop)

• When the brief marks feel load-bearing, autopilot carries the brief's substantive design direction (target vibe, references, principles, the intended feel, design acceptance criteria) into the build-loop invocation, and expects build-loop to run its mandatory, multi-pass visual design loop — render → critique → fix → re-render, every iteration, never a unit-test substitute, never one-and-done.
• In the hand-off, surface the design residual + the human taste spot-check — the loop drives design hard toward the bar, but "looks good to two models" ≠ "a designer signed off." Never imply the loop reached finished/elite design on its own.
• When feel is not load-bearing (a plain utility), plain-but-clear is correct; the visual loop stays light.

The kill-ledger (LOCKED)

Parked or killed concepts are never silently dropped. Each is surfaced with its steelman (the strongest honest case for it) + the specific evidence that would flip it to go — a reviewable parking lot, so a kernel autopilot parked can be rescued. (Maps onto ideate's steelman-before-kill / park-with-a-revisit-trigger.)

The kickoff (emitted each run)

autopilot opens each run by emitting a kickoff that locks the run's frame, then executes it. The full fill-in template is references/kickoff-skeleton.md; it captures: the grounded persona (real lived pain / real data — not invented demand), the forced success-metric + kill-criterion, the autonomy contract, the pipeline order, execute-discipline, the substantive design direction when feel is load-bearing, a narrate-the-phases instruction, and the context/handoff safety-net.

Context / handoff safety-net

A full autonomous run is long and may outlive one chat. Files are the durable memory — the CONCEPT_BRIEF, the pack, build-loop's per-iteration ledger, any proof notes — never chat memory. Each phase's output is a file the next phase (or a future session) reads. If the run risks a context limit, emit a prompt-pack Mode C handoff so a fresh chat resumes cleanly with no amnesia.

Model tiering (spend reasoning where judgment lives)

A long run accumulates paperwork — ledger and receipt copying, status/doc reconciliation, formatting gates, mechanical restatement of a contract already locked in a file. That work does not need frontier-tier reasoning; route it to a cheaper tier (or a script) and keep the strong tier where the run's quality actually comes from.

• Strong tier: the brief's decisions, the design direction, deep-dive's red-team, build-loop's critique passes, and any debugging where the cause is still unknown.
• Cheap tier: transcription, reformatting, filling a settled template, restating a locked decision, mechanical doc sync.
• The hard line: anything making a judgment call about correctness or honesty stays on the strong tier — grading a gate, splitting verified vs. still-a-bet, writing the kill-ledger or the hand-off. Cheap-tiering a judgment call is how a run starts quietly overclaiming.

Pitfalls to avoid

• Re-implementing a sub-skill instead of invoking it. Compose, don't copy. (If you're pasting a funnel or a loop procedure, stop.)
• Faking or auto-passing a human-only gate in-character — the experiment's sharpest failure. Emit it; never narrate a gate you abandoned.
• Treating a synthetic persona's reaction to the solution as validation — a firewall breach. Real data discovers + seeds; it never validates the solution.
• Claiming the output is finished / elite / validated. It's a first draft; the market is the handoff; design has a human taste residual.
• Silently dropping a parked concept — the kill-ledger forbids it.
• Opening a unit's build before its gate-impact pass — discovering constraints one at a time turns one unit into a chain of stop-fix-resume detours, and each partial fix breeds the next.
• Letting the build phase go one-and-done or unit-test-substitute — build-loop's own rules require a genuine multi-pass loop and (when load-bearing) the visual design loop; surface the iteration count + design residual, don't paper over them.

Scale heuristics

autopilot is the autonomous tier of the idea-to-ship suite: it flies the manual tier's pipeline to a near-finish-line-aimed first draft + an honest ledger — eyes open on the 3 limits, never overselling the result.