| name | omd-autopilot |
| description | One-prompt autonomous product design and implementation. Use automatically for broad greenfield UI requests such as 'from scratch', '새 제품/화면을 알아서 만들어줘', or requests that delegate DESIGN.md creation. It decides whether to reuse, establish, refresh, or skip a project design system; asks at most one consequential question batch; then builds and verifies the real surface. Use omd:harness instead only when the user explicitly asks for guided checkpoints. |
omd:autopilot
Use this skill for an ordinary natural-language request to create a new product
surface without requiring the user to name a skill, choose a reference, or set
up the guided harness.
This is a separate workflow from omd:harness. Never claim to have approved or
bypassed its mandatory checkpoints.
Ownership
The current main host agent is the only implementation owner. Specialists are
read-only advisers. They may write only inside the current run's advisory
folders and must never edit DESIGN.md or product files.
State machine
-
DETECT — inspect the repository, real consumer route, stack, existing
DESIGN.md, preferences, components, states, assets and protected behavior.
When OMD_AUTHORITY_CONTROLLER_RUN_DIR is present, use that exact relative
path as the one run directory; do not derive or substitute a slug.
Create the run-scoped task.md while retaining the exact user-prompt bytes
verbatim (extra clearly labelled code observations may follow, but may never
be presented as user authority), then run
autopilot-mission.cjs <project-root> <run-dir> bootstrap. This freezes the
initial product tree and mission budgets before any product write.
-
AUTHORITY_GATE — run scripts/design-council-prime.cjs in the run scope.
Freeze its decision ledger before any product write.
-
BOUNDED_COUNCIL — dispatch no more than three evidence-required, read-only
lanes. Do not dispatch a lane for a settled decision. A generic authorized
greenfield mission uses only the design-system and interaction lanes;
locale/copy or explicit external-evidence needs may add one relevant third
lane. After the authority
handoff reaches PROPOSE_PLAN, run
autopilot-council-plan.cjs <project-root> <run-dir>, dispatch exactly the
listed roles exactly once and in parallel (in bounded external-controller
mode, execute those same lanes inline per the budget section instead of
spawning advisers), then collect each result once and run
autopilot-council-reconcile.cjs <project-root> <run-dir>. The reconciled
receipt is mandatory and never grants product-write authority. Every lane
must write the exact JSON shape declared in plan.json. Never send a
follow-up, retry, or reformat request for a malformed/missing adviser result;
fail the council honestly and preserve implementation time instead.
-
CONSEQUENTIAL_INTERVIEW — ask zero or one batch. Ask only unresolved
product-authority decisions that materially change acceptance or the design
system. A sufficiently authorized prompt proceeds without a question. Never
create, infer, or edit council-intake.answers.json on the user's behalf;
that file may contain only an actual user response relayed verbatim after
the controller has entered CONSEQUENTIAL_INTERVIEW.
-
DESIGN_SYSTEM_DISPOSITION — resolve exactly one of reuse, establish,
refresh, or surface-local-only. A missing exact brand source is blocked.
After the council handoff reaches PROPOSE_PLAN, run the installed
design-system-plan.cjs <project-root> <run-dir> helper. Its
design-system-decision.json receipt is mandatory before any product write.
-
SYSTEM_PROOF — for establish or refresh, use the contract in
references/design-system-contract.md. The design-system architect may
propose; the main agent writes only run-scoped graph/provenance/coverage
drafts. New generation, synthesis, refresh, and refactor are single-write
Core v2: never emit legacy frontmatter or 13/15/16-section layouts. After the
design-system-decision.json receipt grants establish/refresh authority,
inspect the environment before preparing a review. When
OMD_AUTHORITY_CONTROLLER_RECEIPT is present, the main agent is explicitly
not the project owner: it must not run either approval helper, pass a
--reviewer, assert --authority-transition-approved, calculate a hash, or
choose a second output name. Author the three drafts once, with every
interactive component declaring all seven state-applicability entries and
every non-interactive component declaring only a reason (non-interactive
error/success display variants do not require a focus-visible state).
Before spending the single activation, validate the drafts with the
controller's provider-free dry-check. It may be run any number of times and
never counts against the activation budget:
node $OMD_AUTHORITY_CONTROLLER_EXECUTABLE --dry-check . $OMD_AUTHORITY_CONTROLLER_RUN_DIR
The dry-check compiles the drafts into a scratch package and verifies that
every evidence path referenced by provenance.json and coverage.json
(for example council/<lane>/result.json) exists as a real file at the
project root. Fix every reported issue and rerun the dry-check until it
prints "status": "dry-check-pass". Only then invoke exactly once:
node $OMD_AUTHORITY_CONTROLLER_EXECUTABLE . $OMD_AUTHORITY_CONTROLLER_RUN_DIR
Both commands must be issued standalone, byte-exact as written — never
append ;, &&, echo, redirects, or any other text to either command.
That provider-free helper binds the preregistered external controller,
compiles from the prepared review's normalized inputs, creates the exact
checkpoint, adopts atomically, and runs project validation. If it fails,
preserve the single failure and stop system work—never create review-v2,
package-v2, or a replacement mission. This path exists to protect the
product-build budget; after success, move directly to the acceptance plan
and real route, giving the explicit unavailable-information state the same
implementation priority as default/focus-visible.
-
ACCEPTANCE_PLAN — before product admission, materialize
acceptance-plan.json. Quote the exact task bytes for every journey,
constraint, and protected unknown. Lock the real route, default/loading/
empty/error/success/disabled states, 1440/390/320/200%-reflow viewports,
and the exact functionality/journey/responsive/keyboard/accessibility/
honesty/design-conformance checks. A generic checklist is not admission.
Preserve every positive journey and supported-item claim at equal or
stronger semantics. An honest unavailable, unknown, deferred, or fallback
state may coexist with a required journey, but it never satisfies or
replaces that journey unless the prompt explicitly makes that exact item
unavailable. For example, “start a reservation” requires a newly operable
reservation-start state, not only a notice that reservations are
unavailable; a stated five-locale surface requires localized core content
in all five locales even when a secondary translation resource has an
unavailable state. Reject the plan and revise it before product admission
when one requirement weakens or contradicts another.
-
PRODUCT_BUILD — implement the requested real route and all required
empty/loading/error/success/disabled states. Apply only proven or explicitly
proposed project tokens. When a controller execution budget is present,
finish authority and council work before its handoff reserve begins. The
reserve belongs to implementation, acceptance proof, and controller
handoff—not additional research or adviser repair. Treat zero document
overflow at 390px, 320px, and 200%-reflow as a product requirement, and keep
primary task controls at least 44×44 CSS px on touch viewports unless the
control is an inline prose link or a native control whose associated label
supplies the target. Treat state transitions as product contracts: a
validation error moves focus to the failing control and is programmatically
associated with it; a success status names the affected record/action and
remains reflected in the source collection or detail state. Run contrast
checks on enabled and disabled task controls—not only the final success
state. A filterable collection must retain a meaningful baseline dataset
that makes the filter outcome observable; when a native select is used, its
selected option is both the programmatic and visible active state. If a
progressbar role is present, keep aria-valuenow and aria-valuemax
synchronized with the visible progress text in every state and locale.
Never hide focusable descendants with aria-hidden alone: use hidden,
inert, or remove/disable their focusability until the state opens. When an
acceptance requirement makes an honesty boundary observable (for example
fictional sample data or “not medical advice”), render that boundary as
visible accessible product copy rather than keeping it only in source notes.
-
VERIFY — verify functionality, same-route desktop/mobile/320px/200%,
keyboard, accessibility, responsive behavior, copy, evidence honesty and
DESIGN.md-to-code conformance. proof.json schema 0.2 must bind the mission,
acceptance plan, product-build admission, route, exact current product-tree
SHA, repair round, every task requirement, and every quality check. Each
atomic result needs non-empty evidence. pass is the conjunction computed
from those results; prose confidence or a self-authored summary is not proof.
“Browser unavailable”, skipped checks, or missing screenshots must be a
failed check, never a passing substitute.
When .benchmark/controller-verification-policy.json is present, the
installed mission controller is the objective-verification authority. Every
local proof, passing or failing, must stop at EXTERNAL_VERIFY before any
repair budget is consumed; do not write delivery, start a local repair, end
the mission, remove the policy, or invent its receipt. The host controller
evaluates the real route and supplies the next hash-bound state. If the
controller passes while a broader local check still fails, the remaining
local failure may then use the same bounded repair budget.
In this controller-owned mode, never discover, install, launch, or probe a
local browser, Playwright/Chromium binary, HTTP server, screenshot command,
browser harness, or GUI application. Do not spend the controller handoff
reserve testing whether those tools exist. Finish deterministic source
checks, write the truthful proof, advance to EXTERNAL_VERIFY, and return
control immediately; the controller owns all browser execution.
-
BOUNDED_REVISION — the main agent may apply at most two focused repair
rounds in the same mission. The controller writes an exclusive receipt for
every failed proof, freezes the exact failed requirement/check IDs, and
requires both a changed product tree and a replacement proof at the next
round. Critics stay read-only. Unresolved BLOCK produces a failed handoff.
A controller-authorized round is an internal continuation of the same
one-prompt mission, not a retry or replacement. Read only its exact
.benchmark/controller-feedback/round-<n>.json, preserve passing behavior,
update the product and atomic proof for that round, and return to
EXTERNAL_VERIFY. Never bootstrap a second mission to escape the findings.
-
HANDOFF — report implemented files, system disposition, question count,
proof hashes, screenshots, failures, time and token coverage.
Run autopilot-mission.cjs <project-root> <run-dir> advance at every state
boundary. The controller rejects product edits before authority, limits
pre-proof project changes to the exact compiler-produced adopted package, issues the product-build admission only after
an exact system proof and acceptance plan, recomputes atomic proof pass, and
refuses to force-pass or replay an exhausted repair budget.
Only one project-scoped Autopilot mission may be active. Continue its bounded
repair loop in the same run; never create a second run to replace, retry, or
escape an unresolved active mission. Completed and failed missions are
terminal and non-resumable.
Bounded external-controller budget
When OMD_AUTHORITY_CONTROLLER_RUN_DIR is present, the mission runs under a
hard wall-clock budget and the product route is the graded deliverable. The
design-system rigor stays intact — what changes is where the minutes go.
- Check elapsed time (
date +%s) at every state boundary. Authority, council,
and system work together must finish inside the FIRST 40% of the budget;
everything after belongs to PRODUCT_BUILD → VERIFY → controller handoff.
- Dispatch NO adviser subagents in this mode. Execute the planned council
lanes inline: the main agent authors each lane's exact JSON shape as its own
read-only analysis from DETECT evidence. Inline lanes still never grant
product-write authority; reconcile normally. A subagent round-trip you can
answer yourself from the repository is budget theft from the product.
- Author the three system drafts in one pass, run the controller
--dry-check
until it passes, then invoke the controller once, immediately. Do not
re-read, re-verify, or beautify drafts the compiler will normalize anyway —
the dry-check IS the verification step, and it is free (it never counts
against the single activation).
- The dry-check catches schema/enum violations and missing evidence files
(every path referenced by provenance/coverage evidence must exist at the
project root). A guessed enum or phantom evidence path wastes the single
activation — the controller fail-closes on the first violation and there is
no second invocation, so never invoke the real activation while a dry-check
is still failing.
- The controller invocation must be the ENTIRE command — never append
;,
&&, echo, date, or anything else to it (sequencing voids the
exactly-once contract). Run elapsed-time checks as their own separate
commands before or after.
- Before invoking the controller, ensure the exact
task.md bytes also exist
at the project root (copy from the run directory): the adopter requires
<project-root>/task.md as proof evidence.
- If authority+system work has consumed 50% of the budget before adoption,
skip every remaining optional analysis and go straight to the smallest
compiler-valid drafts.
Product route order (graded-state first)
In PRODUCT_BUILD, implement in this exact order — a polished page missing a
required state scores zero, an honest skeleton with every state scores:
- Semantic skeleton for the real route: landmarks, single
nav with
disclosure collapse, skip link targeting #main (never the primary CTA),
heading order, form field ID graph (label[for], aria-describedby
hint+error chain, role="alert" errors, focusable role="status" success).
- EVERY required state as a real interaction outcome, before any visual
polish. The task brief's journey verbs ARE the graded states. Rewrite
the brief as a verb list first ("filter X", "inspect one Y", "assign a Z",
"switch locale", "mark progress", "see completion") — each verb is one
state that must be REACHABLE by real keyboard interaction (Tab to the
control, then Enter/Space) on the product's own controls. A state that
cannot be reached by the described action does not exist — and a state
reached only through a developer switcher does not exist either: NEVER
render "Show " radios, demo toggles, or any state menu in the
product UI.
data-state="<state-name>" markers live on the real
components that enter those states. Every entry in the system's
honesty/unknown ledger renders as a visible unavailable-information
node — prose disclaimers alone do not count.
- Programmatic-semantics parity — every state change updates the visible UI
AND the machine contract in the same paint:
- Selection/toggle:
aria-selected / aria-pressed / aria-checked /
aria-current on the control itself, and for filters also a visible
role="status" or aria-live summary ("Showing only urgent …").
- Validation error: focus returns to the offending field, the field gets
aria-describedby pointing at a visible role="alert" (or
aria-live) message — all three together, not any one alone.
- Persisted outcome (e.g. an assignment): announce it in
role="status"
WITH the record's visible ID and the chosen value, and update the source
record's own text to show the same value.
- Locale switch: update
<html lang>, the selector's committed value, and
the rendered script together; never silently change the selected
language. Progress: role="progressbar" aria-valuenow/max must equal
the visible "N of M" text.
- Detail surfaces are
role="dialog"|"region"|"complementary" containing
the record's visible ID; records carry stable visible IDs.
- Name controls with the brief's exact nouns and verbs ("Reserve a tool",
"Assign owner", "Urgent") — the brief's language is the accessible name,
verbatim, with no decorative words prepended or appended.
Before declaring the product finished, run a SELF-WALK: list every journey
verb from the brief, and for each write (a) the exact keyboard path that
reaches it and (b) the programmatic evidence that proves it (which attribute
or role changes). Any row missing either entry is unfinished work — fix it
before the mission proof, budget permitting, because a missing row scores the
same as a missing page.
Philosophy derivation chain (mandatory order)
Read references/derivation-chain.md before any system draft. The order is
PHILOSOPHY → DERIVE (decision table with D-ids and rationales) → TOKENS
(with D-id back-references in comments) → COMPONENT SPECS (documented before
code; select from references/presets/INDEX.md first and derive token slots
from the decision table — never improvise from zero what the preset catalog
already validates (gate GS8); references/component-craft.md is the floor)
→ LAYOUT GRAMMAR (per page, with content back-calculation — if the data
cannot fill the grammar, enrich the data first, never leave wide viewports
empty) → BUILD (pages consume only) → RENDER CRITIQUE → DESIGN.md carrying
the philosophy and decision table so a designer can read WHY every value is
what it is. A token value without a D-id rationale is an improvised value
(gate GS7).
Data discovery (good UI carries the data)
A screen is only as good as the data it carries. Before designing anything,
inventory the DATA TRUTH SOURCES and write data-inventory.md into the run
directory:
- Provided fixtures first:
data/*.json, data/*.js, CSV, seeded stores.
- Declared contracts next: OpenAPI/Swagger specs, GraphQL schemas, TypeScript
interfaces/types, ORM models, API route handlers.
- If neither exists and the brief implies records, the mode decides:
- Guided mode: ASK the user for the data shape, or ask permission to scan
the repository for data sources before proceeding.
- Autonomous/benchmark mode: use only what was provided; state the absence
honestly on-screen. NEVER invent fields, records, or endpoints.
Then design FROM the data shape, not toward a template:
- Entities and their cardinalities pick the page patterns (a 34-row entity is
a table with filters, not three cards) and the density dial.
- Every enum in the data (status, category, stock) becomes a system token
set: one color/mark per value, defined once, legible at a glance.
- Every field a user needs is on screen or one interaction away; IDs and
timestamps are visible where an operator would search by them. A field in
the dataset that never renders anywhere needs a reason.
- Aggregates shown as KPIs are COMPUTED from the data at runtime and their
definition is stated near the number (e.g. "open = pending + packed").
- When an API is the source (now or planned): the schema is the contract —
per-endpoint loading/error/empty states, pagination beyond ~50 rows,
and no rendering of fields the contract does not define.
Visual quality contract (mandatory read)
Before PRODUCT_BUILD, read references/visual-quality-contract.md in full
and treat every item as a gate, not a suggestion. Non-negotiables repeated
here because violating them wastes the whole run:
- NEVER render developer/state switchers in the product UI; states come from
real product interactions only.
- NEVER ship native unstyled form controls — restyle every control from
system tokens with the accessible native input underneath.
- The system carries a display/body type pair with a ≥2× display step, a
label role, section-air spacing steps (2.5–6rem), ONE surface genre, and
accent as a small signal (≤~5% of any viewport).
- After building, run the ONE-round self-critique from the contract (5 axes
- gate sweep, System Fidelity replaces variety) and write
critique.md
into the run directory before the mission proof.
Multi-page products
When the brief asks for more than one page, the design system is the
consistency contract: ONE shared stylesheet owns tokens and components
(defined exactly once); every page links it and adds only page-level layout.
Nav and footer are designed once and rendered identically on every page with
aria-current on the active link. Body, heading, and primary-action styling
must compute identically across pages; every internal link resolves to a
real page. Build page one as the system's proof, then express the remaining
pages FROM it — if page two needs a new token or component, that is a system
change first, not a page-local invention.
Motion and micro-transitions (system-owned)
Motion is a design-system concern, not per-element improvisation. When the
brief asks for a modern/polished feel (or names transitions), establish
motion TOKENS in the system first and cite only them:
- One duration scale (fast ~120ms, base ~200ms, slow ~320ms) and at most two
easing curves (one enter, one exit). Every transition on the page uses a
token pair — a one-off cubic-bezier is an unauthorized token.
- Animate ONLY
transform and opacity (compositor-friendly); never animate
layout properties (width/height/top/margin) or box-shadow directly (fake
elevation with a pseudo-element opacity fade).
- Standard vocabulary, applied sparingly: entrance fade-up for major sections
(one-time, on first reveal), hover elevation/tilt on interactive cards,
pressed-state scale on buttons, focus-visible transition on CTAs. One page
needs 3–5 total motion patterns, not one per element.
- EVERY animation sits behind
@media (prefers-reduced-motion: reduce) with
a non-animated equivalent state — reduced-motion is a graded contract, not
an afterthought.
- Provided asset images get explicit width/height (no CLS) and motion applies
to their container, never the raw img.
Framework idiom projection
The adopted design system is stack-neutral; the PRODUCT expresses it in the
stack the brief names. Detect the stack from the brief and any provided
runtime (vendored libraries in assets/), then project tokens and state
semantics into that stack's native idiom — never a foreign one:
- Vanilla HTML/CSS: tokens as
:root custom properties; states as
data-/aria- attributes driven by small event handlers.
- React (including buildless Preact+HTM): tokens as one exported theme
object AND mirrored
:root custom properties; components as functions
whose props carry state; aria attributes computed from the same props that
drive the visuals (single source of truth — never a DOM query after
render); lists keyed by stable record IDs; no innerHTML string templating.
- Tailwind (when present): tokens declared once in the config/theme layer
and referenced by utility classes; never inline arbitrary values that
bypass the token scale.
- The same journey states, aria semantics, and honesty rules apply
identically in every stack — the projection changes syntax, never the
contract.
Design-system decision
- Valid compatible root
DESIGN.md → reuse without reopening it. Legacy
13/15/16-section and unmarked documents remain readable during the compatibility
window; reusing one does not silently rewrite it.
- Explicit or delegated authority to build a system →
establish.
- Explicit replacement of an existing system →
refresh; legacy input must pass
staged migration/check and opaque-extension preservation before replacement.
- Narrow repair or explicit refusal →
surface-local-only; never promote local
choices as project facts.
- Broad greenfield with no authority → ask one question: project system or
local surface contract.
- Exact official brand request with no authoritative source → block rather
than fabricate.
Reference selection happens only after this decision and only when it supplies
useful verified inspiration. A reference never owns product facts.
Evidence and unknowns
Classify each consequential system decision as prompt-fact,
repository-fact, verified-reference-inspiration,
agent-proposed-greenfield-decision, or unresolved.
Unknown means absent at the smallest boundary. Never synthesize a company fact,
font, component, metric, testimonial, price, security promise, or narrative.
Core v2 does not require placeholder facts: omit unresolved values from tokens,
prescriptive prose, and code. A consequential unresolved decision may be named in
Governance without a suggested fallback.
Required run artifacts
Store permanent artifacts under .omd/runs/<run-id>/:
mission.json
council/decision-ledger.json
design-system-decision.json
system/proposal.md, migration report/rollback references when applicable,
and generated system/proof.json
implementation.json
acceptance-plan.json
proof.json, repairs/round-<n>.json, and screenshots
delivery.json
Receipts bind the original task, repository evidence, DESIGN.md, product output,
consumer route, states, viewports and validator results. Missing proof is not a
pass.
The project-owned canonical system lives outside the run at:
.omd/system/manifest.json
.omd/system/graph.json
.omd/system/provenance.json
.omd/system/coverage.json
The visible DESIGN.md begins with # <Product> Design System, contains exactly
the seven neutral design-md:section anchors in the frozen Core order, and has no
YAML/frontmatter, OmD/tool/generator/quality metadata at its top. Only an adopted,
valid profile: portable-core manifest with exact graph/projection hashes makes
the graph canonical; a migration candidate keeps its named source DESIGN.md
canonical. The seven semantic design-md:claim declarations and every
design-md:claim-end are compiler-owned and must not be edited after rendering.
The Markdown remains a complete portable contract on its own.
Guided-mode boundary
If the user explicitly asks to review journey/system/validation checkpoints or
to collaborate phase by phase, route to omd:harness and preserve all of its
mandatory checkpoints. Do not silently switch an active guided run to
Autopilot.