Prepare a task — turn a free-text backlog item into a well-formed todo. Take backlog → pending → todo, ask clarifying questions, set acceptance anchors (Automated / Manual split), do the Figma design-system-first split, classify builders and builder order, detect blockers and dependencies. Triggers on "prepare a task", "prep TASK_12", "backlog to todo", "task questions", "acceptance anchors", "split a task", "classify builders", "promote a backlog item".
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
Prepare a task — turn a free-text backlog item into a well-formed todo. Take backlog → pending → todo, ask clarifying questions, set acceptance anchors (Automated / Manual split), do the Figma design-system-first split, classify builders and builder order, detect blockers and dependencies. Triggers on "prepare a task", "prep TASK_12", "backlog to todo", "task questions", "acceptance anchors", "split a task", "classify builders", "promote a backlog item".
Owns the task-intake + task-prep work. This is the
operational entrypoint between a free-text user idea and a structured task the
orchestrator can execute; it produces questions when scope is unclear and a
promoted todo when it is clear. It never writes product code. Long normative
rules live in this skill's references + frozen contracts (routed below).
When to use
Promote a fresh backlog file (Mode A) — prep TASK_<N>_<title>.
Re-evaluate an answered questions sidecar (Mode B) — the user filled ### Answer blocks.
Classify a task against the change-kind taxonomy and emit its builder order.
Not this skill: writing Kotlin/JS/product code, invoking product builders or
product validators (orchestrator's job, only after the task is in todo/),
editing done/. The canonical task-state validator and transition helper are
mandatory publication infrastructure, not product validation.
Required inputs
Mode A — orchestrator/tasks/backlog/TASK_<N>_<title>.md with the required first line # TASK <N> — <title>.
Mode B — the answered orchestrator/tasks/pending/TASK_<N>_<title>.questions.md (path or inline form; before analysis, helper-persist inline bytes verbatim, re-read and prove byte-for-byte/SHA-256 equality, then obtain a fresh canonical sourceRevision).
The caller's prompt states which mode. <STEM> = the filename without .md / .questions.md.
Board Prepare no-questions contract
When the caller includes the exact marker BOARD PREP POLICY: NO QUESTIONS.,
apply this action-scoped contract to that run only. It belongs to the
server-owned Board Prepare action; it is not a project/global setting and
does not change direct task-prep invocations.
Never emit needs_action, invoke the ask transition, or publish a
pending/*.questions.md sidecar.
Resolve repo-decidable gaps from evidence. For a reversible ambiguity, choose
the safest conservative default and record it as an Assumed — Input.
For irreversible, destructive, authorization, breaking-contract, or
missing-owner ambiguity, keep the unsafe portion out of scope and return a
typed actionable blocker/follow-up. Do not turn it into a user question.
If a safe runnable scope remains, compose it and promote to todo. A typed
blocker is preferable to guessing, but it is never a request for an answer.
Workflow
Pre-flight + authority + lock — verify the required references and select exactly one guarded writer authority: inherited site session, exact caller-owned standby lease receipt (verify it; never reacquire or release it; renew only as specified in references/prep-flow.md), or a newly acquired prep-owned direct lease. Run the edit-marker guard. If the caller supplied a complete task-lock receipt, verify its exact stem/stage/runId/sessionId/lockHash; otherwise acquire one no-clobber generation through task-lock.mjs. A standby acquire passes the caller receipt's exact sessionId, leaseId, and private token; the helper verifies that capability before and after publication without persisting or returning the token. Retain the complete lock receipt privately. For Mode B inline, first validate pending state, helper-persist the prompt's exact answer bytes through persist-answers --input -, require the matching receipt, and re-read exact byte/SHA-256 equality; then re-run canonical validation and retain its fresh revision for analysis and any second transition. persist-answers permits only Answer-body changes in the current round and rejects answer markup that changes the CommonMark-visible question/Answer heading identity; ask is reserved for a newly generated round and enforces the generation counters. Never analyze inline answers before that durable first transaction, and never reuse its pre-persistence revision. For every other input, validate the exact current stem plus fresh INDEX and retain the mutation sourceRevision. Emit phase-start. Never write or delete lock bytes directly.
Classify (Step 1) — reuse the change-kind taxonomy from references/intake-classification.md verbatim (do not re-derive); record the builder kinds implied + which prerequisite scaffolds exist.
Builder order — order builders low-layer → high per orchestrator/contracts/builder-order.md: toolkit/compose-lib FIRST, then data-service-scaffold → data-feature → ui-core-state → mapper → feature-module-scaffold → screen/dialog → cross-feature-nav → resource → app-shell LAST. Prepend missing prerequisites (feature-module-scaffold before screen, ui-core-state before mapper/screen, data-service-scaffold before endpoint/).
Stop and ask (BLOCKED / ESCALATE)
Backlog missing the # TASK <N> — <title> first line → BLOCKED (don't auto-number).
An ## Outcome appendix present in backlog/pending/todo → BLOCKED; preserve it and reconcile through canonical finalization/reopen recovery. Never strip Outcome by hand.
Stem already lives in another column (todo/ or done/) → BLOCKED (collision; a stem is in exactly one column).
A promoted ### Automated bullet with no automation anchor → BLOCKED (don't silently move to Manual).
A self-dependency or dependency cycle → BLOCKED. An ordinary dependency missing
from done/ stays in ## Depends on; preparation may promote it and Run will
remain closed until the prerequisite is accepted in done/.
Convergence stall (gap count not shrinking two rounds running) → ESCALATE (split / rewrite / drop).
References to read
Self-contained reference pack — the prep MECHANICS live in references/. Read
only these references + the frozen contracts at runtime; this skill reads no
external rule docs. Route via references/index.md.
topic
read
Routing table + file map (start here)
references/index.md
Change-kind taxonomy + prerequisites + builder order + validator gate + intake plan
references/intake-classification.md
Full prep flow (Modes A/B, pre-flight, lock, journal, gap analysis, questions, Mode B convergence, canonical INDEX publication, output blocks)
Cross-reference (owned by another skill — see references/index.md):
orchestrator/tasks/README.md (lifecycle/board),
../validation-gates/references/when-to-stop-and-ask.md (stop-and-ask rules).
The cookbook, project-config, and figma screen-cache/census rules live in their
owning skills' references/**.
Always pre-flight every file with [ -f <path> ] before reading; a miss is a BLOCKED return.
Validators / gates
No product-code validators — prep runs before the orchestrator pipeline.
Every durable task mutation nevertheless passes the canonical task-state
validator and deterministic transition helper. The semantic prep gates are:
Source provenance gate — a real canonical ## Source immediately after
the H1 is immutable. Copy its exact block into every todo proposal; never
derive it from prose, rewrite it, or put it in a pending questions sidecar.
When Source Type is api-work-package, the canonical
## API Work Package section is part of that immutable provenance: preserve
the complete section byte-for-byte immediately after Source in every todo
proposal.
Convergence cap — Mode B iterates while the gap count strictly shrinks; two
consecutive stuck rounds ESCALATE (T_now >= T_prev and T_prev >= T_prev_prev).
7-questions-per-round UX cap (render limit only; not the gap total).
Dependency gate — preserve unresolved dependencies during promotion;
block only self-dependencies/cycles here. Canonical Run admission requires
every dependency to be accepted in done/.
Cross-column collision gate — a stem must exist in exactly one column.
Acceptance-anchor gate — every ### Automated bullet carries a file/class/gradle anchor.
Cache pre-flight gate — the only hard stop in the Figma split path (non-none designs).
Output contract
Return the single Step-8 Markdown block matching what happened — Mode A promoted /
questions, Mode B promoted-after-answers / more-questions, or ESCALATE (see
references/prep-flow.md for the exact shapes) and the frozen
task-prep contract. Builder
ordering in the plan MUST follow orchestrator/contracts/builder-order.md. On-disk
products: the helper-published pending/<stem>.questions.md sidecar (ASK path)
or todo/<stem>.md task (PROMOTE path), plus the helper's verified fresh
INDEX.json. Every promoted task body preserves the source task's canonical
## Source block byte-for-byte and, for Source Type api-work-package, its
complete canonical ## API Work Package section byte-for-byte immediately
after Source; ## Origin remains separate lineage.
Task-prep keeps proposal bytes in memory and delivers them only
through the helper's bounded --input - stdin contract; it never mutates
durable column files or stages proposal artifacts itself.
room-migration
Gap analysis (Step 2) — spawn context-finder + requirements-lookup in parallel; build the gap list (missing inputs, unbounded scope, multi-feature, resource-only, migration triggers, contract-missing endpoints, missing screen designs).
Decide (Step 3) — split gaps into decidable (repo/convention/precedent answers it) vs undecidable (product intent, irreversible migration/destructive/authorization, ambiguous owner, missing artifact). Draft a grounded default for every decidable gap (record it as an Assumed — Input) and PROMOTE; ASK only the undecidable ones. Under BOARD PREP POLICY: NO QUESTIONS., ASK is forbidden: use the action-scoped safe-default/blocker rules above and promote whenever a safe runnable scope remains. Bias toward drafting, not asking.
Ask (Step 4) — this step is unreachable under BOARD PREP POLICY: NO QUESTIONS. Otherwise, ask only undecidable gaps, and every question proposes a concrete default to confirm (**Recommended**: for choice, **Proposed**: for text) — never blank. Compose the complete questions sidecar in memory, then pipe its exact UTF-8 bytes to transition-task-state.mjs ask --input - with the exact fresh revision and writer authority. Close stdin and require its v1 pending-state receipt; never write the durable sidecar directly or write an intermediate proposal file.
Compose promotion (Step 5) — compose the standard todo shape in memory, including acceptance anchors and defensive boundaries, and freeze its proposed ## Design section. Do not invoke promote until Step 5.5 has finished and every promotion-time augmentation is folded in.
Figma census split + promote (Step 5.5) — gated on figmaEnabled: true + any non-none## Design bullet: pipe the proposed parent bytes to the cache pre-flight through FIGMA_SCREEN_TASK_FILE=-, run component census, AMBIGUOUS/RETIRED/SOURCE_STALE/UNSUPPORTED → ask (owner registry/inventory remedies) in the normal direct flow, but return the exact typed remedy as a non-question blocker under BOARD PREP POLICY: NO QUESTIONS.; MISSING/INCOMPLETE → create each child through the canonical deterministic absent→backlog creator and then promote it through the same transition helper; never mint a number or write absent→todo. Chain the validated child stems in the parent, append preview/spec/screenshot acceptance, then pipe the final exact UTF-8 bytes to transition-task-state.mjs promote --input -. Require its v1 todo-state receipt. The helper alone writes todo, detaches backlog/pending, validates, rolls back, and publishes INDEX.
Mode B convergence (Step 6) — classify each answer (clear/partial/off-target/empty/SKIP), re-run gap analysis, compute T_now; PROMOTE at 0, ASK on strict shrinkage, ESCALATE on two stuck rounds.
Verify receipt + release — verify the helper's final postcondition/fresh-INDEX receipt and re-read the durable result. Release only the exact lock generation through task-lock.mjs release with the retained run/session/hash tuple and that final receipt's exact state/sourceRevision; on a non-transitioning BLOCKED/ESCALATE exit, first obtain an equivalent fresh green state + INDEX receipt. Require the release receipt to echo those values and a valid snapshot hash. If its atomic detach was interrupted, use idempotent recover-release with the same lock tuple plus a fresh final state/revision, and never delete .release-* by hand. Then release only a writer lease that task-prep itself acquired; a caller-owned standby lease is released by standby after the full prompt returns. A missing/mismatched lock or stale postcondition is a recovery condition, never a raw-delete no-op. Never run a second INDEX publisher. Emit phase-end / stop and return the Step-8 block.
