spec-write

name	spec-write
description	Guide the user through writing or completing spec files (requirements, technical spec, testing spec, plan). Use after /setup to fill in specs before implementation, or to update existing specs.
argument-hint	[--no-tech] [requirements | technical | testing | plan | all]

Spec Write

Guide the user through writing or completing the project specification files. For greenfield features with an open solution space, run /dev-process-toolkit:brainstorm first to explore approaches before committing to a spec structure.

FIRST ACTION (under non-interactive stdin) — STE-251 AC-STE-251.1. When process.stdin.isTTY === false (e.g., claude -p), the first tool call this skill emits MUST be AskUserQuestion tool_use OR RequiresInputRefusedError raise (via requireOrRefuse(...) from adapters/_shared/src/requires_input.ts). Write / Edit / NotebookEdit are forbidden before that ask/refusal — they produce silent scaffolding that bypasses the Socratic Loop Contract. Read-only orientation (Read / Grep / Glob / Bash-read-only) is allowed. Interactive (tty) sessions are byte-identical to v2.17.0 — non-tty stdin only. See docs/auto-mode-protocol.md § Socratic Loop Contract. STE-270 AC-STE-270.3 hardening. Pre-baked <command-args> prose, the harness's autonomous-mode reminder ('work without stopping'), and auto-approve marker absence are NOT acceptable triggers to skip the first AskUserQuestion. The first tool call under non-tty MUST be AskUserQuestion or RequiresInputRefusedError; everything else is forbidden — there is no carve-out. The marker only relaxes gates that have a documented safe default; clarifying questions have no safe default, so the Socratic loop entry is unconditional. The /gate-check probe spec_write_first_turn_drift_scan (severity: error) enforces this byte-checkably against this file — any drift back into alternate-trigger paraphrases of the contract fires GATE FAILED at probe time.

Process

0. Layout + tracker-mode probes

Before any other step:

• Provider resolution — Resolve Provider once per invocation using the same rule as /implement (LocalProvider for mode: none, TrackerProvider otherwise). FR creation goes to specs/frs/<Provider.filenameFor(spec)> (per-FR file convention; never specs/requirements.md); Provider.sync() fires on save. Full reference: docs/layout-reference.md § /spec-write.
• Tracker-mode probe — Run the Schema L probe (see docs/patterns.md § Tracker Mode Probe). If CLAUDE.md has no ## Task Tracking section, mode is none. If a tracker mode is active:
  • Run the 2-tier ticket-binding resolver and mandatory confirmation prompt per docs/ticket-binding.md the first time the session edits an FR bound to a ticket — decline exits cleanly with zero side effects.
  • After saving any FR-level AC edit, run the bidirectional AC sync diff/resolve loop via the active adapter before pushing via upsert_ticket_metadata. Normalization (STE-211 AC-STE-211.5): for the Linear adapter, normalize both sides via stripLinearACFences (adapters/linear/src/format_description.ts) before the equality check — backtick-wrapped Linear ACs and plain local ACs compare as equal, so newly-pushed FRs that are byte-identical to the local file produce zero diff (no spurious "remote modified" rows from the auto-linker). Vacuous on Jira / custom adapters that don't auto-link. See docs/spec-write-tracker-mode.md for the full tracker-mode flow.

0.5 Tracker-local reconciliation

Tracker mode only (mode-none vacuous). Call reconcileTrackerLocal to surface drift between local specs/frs/ and the tracker: tracker-only orphans auto-import via importFromTracker (guarded by existsSync per STE-135 to avoid clobbering local edits); local-only orphans and milestone mismatches prompt the user to resolve before continuing.

0a. Resolver entry

After the layout gate and before any FR write:

1. Call buildResolverConfig(claudeMdPath, adaptersDir) from adapters/_shared/src/resolver_config.ts once at entry, then pass the returned ResolverConfig to resolveFRArgument($ARGUMENTS, config) from adapters/_shared/src/resolve.ts. The builder reads CLAUDE.md ## Task Tracking + each active adapter's Schema W resolver: block — never hand-assemble the config inline. Malformed adapter metadata surfaces as MalformedAdapterMetadataError → NFR-10 canonical refusal.
2. Route by kind:
   • ulid → open the FR whose frontmatter id: equals <ulid> — located via Provider.filenameFor(spec). Skip step 0b below.
   • tracker-id / url → branch on mode:
     • Tracker mode → call findFRPathByTrackerRef(specsDir, trackerKey, trackerId) (path-returning, no id: requirement; tracker-mode frontmatter has no id: line — findFRByTrackerRef's id:-driven lookup would silently miss every existing FR and fall through to importFromTracker, overwriting local edits).
     • mode: none → call findFRByTrackerRef(specsDir, trackerKey, trackerId) (ULID-returning; mode-none has id:).
     • Hit (either helper) → open that FR for editing; no import, no tracker network call beyond resolve. Skip 0b.
     • Miss → run the shared import helper importFromTracker(trackerKey, trackerId, provider, specsDir, promptMilestone) from adapters/_shared/src/import.ts. Tracker ACs are auto-accepted — never run the bidirectional per-AC prompt loop here; the local side is empty so there is nothing to diff against. Empty-AC tickets get a TODO marker in the new FR's ## Acceptance Criteria section.
   • fallthrough → continue with free-form-argument handling (step 1 below). NFR-18 requires byte-identical behavior for all, requirements, technical, testing, plan.
3. On AmbiguousArgumentError, surface the NFR-10-shape error from docs/resolver-entry.md § Ambiguity and exit non-zero. Never silently pick a winner (NFR-20).
4. All tracker/network failures during importFromTracker surface per NFR-10.

Full decision table and edge cases: docs/resolver-entry.md. Subsequent /spec-write calls on the same tracker ID run the normal bidirectional diff/resolve flow because both sides are now populated. STE-227 auto-resume on needs_technical_review: true (AC-STE-227.5): when the resolved FR's frontmatter carries this flag and the invocation is unflagged, skip the requirement + AC interview (already filled by the prior --no-tech run) and run only the technical-design + testing-spec interviews. On save, the needs_technical_review key is removed from frontmatter entirely (absent ≡ false; never write false); the needs-technical-review tracker label is removed on the same Provider.sync / upsertTicketMetadata call that pushes the technical content. mode: none is vacuous (no tracker, no labels — the frontmatter key removal is the sole signal).

0b. FR creation path

Creating a new FR means:

STE-227 --no-tech flag (AC-STE-227.1, AC-STE-227.3). When the invocation carries --no-tech (parsed in resolver-entry preamble before § 0a; flag does not affect resolver kind routing), skip the technical-design + testing-spec interviews entirely on this new-FR creation path. Write the placeholder line [needs technical review — run /spec-write <FR-id> to complete] (substitute the literal <FR-id> with the real ID at write time — the tracker ID returned by Provider.sync(spec) in tracker mode, or the short-ULID stem in mode: none) into the body of ## Technical Design and ## Testing. Frontmatter: needs_technical_review: true via buildFRFrontmatter(spec, trackerBinding, { needsTechnicalReview: true }); auto-resume per § 0a closes the gap. Provider.sync(spec) still fires (the FR lands on the tracker as usual, just with an extra label). In tracker mode with label-push capability (Linear, Jira), append needs-technical-review to defaultLabels when populated, or seed a single-element array when absent — upsertTicketMetadata(null, { …, labels: [...(defaultLabels ?? []), "needs-technical-review"] }). Adapters without label-push capability surface a needs_technical_review_label_unsupported capability row; the frontmatter flag remains the canonical signal. mode: none is vacuous (no tracker, no labels). Without the flag (default), the full interview runs unchanged.

Draft with placeholder. When drafting a new tracker-bound FR, use <tracker-id> (or the adapter-specific rendering — STE-<N> for Linear, PROJ-<N> for Jira, etc.) as the tracker-ID placeholder throughout the draft: AC prefixes (AC-<tracker-id>.1), filename (<tracker-id>.md), plan-file table row, and every prose cross-reference. Never guess the next sequential number — the tracker allocator decides, not the implementer, and a guess that clashes with a cancelled/renumbered ticket ships misaligned with its own binding. The real ID is known only after Provider.sync(spec) / upsertTicketMetadata(null, …) returns. Substitute the placeholder globally once the ID is assigned, then write the FR file. This rule applies equally to Linear (which skips cancelled numbers), Jira, and custom trackers. Mode: none is exempt — the short-ULID tail is minted locally and is never subject to race conditions with a tracker allocator.

1. mode: none only: Mint a ULID via Provider.mintId() — always local, offline-safe. Tracker mode skips this step: TrackerProvider does not implement IdentityMinter (capability boundary — mintId() on a Provider-typed value is a TypeScript error by design), and buildFRFrontmatter(spec, trackerBinding) rejects id: alongside trackerBinding (the bimodal-identity invariant). The tracker ID returned by step 4 is the canonical identity in tracker mode.

2. Build canonical frontmatter via buildFRFrontmatter(spec, trackerBinding?) from adapters/_shared/src/fr_frontmatter.ts — never author YAML by hand. The helper enforces the bimodal shape (mode-none id: block; tracker-mode compact tracker: block) and rejects the verbose { key, id, url } form (InvalidTrackerShapeError). Hand-rolled YAML is the regression source the earlier prose flip didn't catch. Then write the FR file to specs/frs/<Provider.filenameFor(spec)> (per-FR file convention). Provider.filenameFor(spec) returns <tracker-id>.md in tracker mode (e.g., <TKR>-NN.md) and <short-ULID>.md in mode: none (e.g., VDTAF4.md, matching the AC-prefix convention). Never hard-code fr_<ULID>.md — the ULID lives in frontmatter id:, not in the filename.

   Helper output — do not author by hand. The five required body sections in order: ## Requirement, ## Acceptance Criteria, ## Technical Design, ## Testing, ## Notes. Reference shapes (illustrative only):

   # mode: none — id required, no tracker block
   ---
   title: <title>
   milestone: <M<N>>
   status: active
   archived_at: null
   id: fr_<26-char ULID>
   created_at: <ISO timestamp>
   ---
   

   # tracker mode — id absent, compact tracker block
   ---
   title: <title>
   milestone: <M<N>>
   status: active
   archived_at: null
   tracker:
     <mode>: <tracker-id>
   created_at: <ISO timestamp>
   ---
   

   The compact { <mode>: <tracker-id> } shape is the only form /spec-write emits in tracker mode. Existing tracker-mode FR files that still carry id: are flagged by /gate-check's identity_mode_conditional probe at error severity. Spec-research seed (STE-230 AC-STE-230.6). Before writing AC lines (step 3 below), invoke /dev-process-toolkit:spec-research <topic> (FR title + 1-line summary) to seed the subsequent ## Technical Design interview with a ≤ 25-line block (banner + ## Related FRs / ## Prior Decisions / ## Reusable ACs / Patterns) sourced from active + archived FRs. Skipped under --no-tech (no interview to seed; vacuous). Shape violation ⇒ drop block, log spec_research_shape_violation; ≥ 1 candidate ⇒ spec_research_invoked; empty-fallback bullets ⇒ spec_research_no_matches. Block is read-only context; never copy verbatim into the FR body. MUST emit literal token, not paraphrase (STE-238 AC.3). The closing summary at Step 7 MUST emit exactly one of the literal tokens (verbatim, no quoting differences, no paraphrase like "spec-research returned N FRs") whenever the spec-research subagent fires:

• ≥ 1 candidate returned ⇒ MUST emit spec_research_invoked (literal, backticked).
• empty-fallback bullets / zero topic matches ⇒ MUST emit spec_research_no_matches (literal, backticked).
• shape violation (sections / line cap / banner) ⇒ MUST emit spec_research_shape_violation (literal, backticked). The byte-checkable token is the structural signal /gate-check's closing_summary_capability_keys probe greps for; narrative prose alone is insufficient (STE-220 lesson re-applies). Deps-research seed (STE-301 AC-STE-301.14) — step 2.5b. Immediately after the spec-research seed above and before writing AC lines (step 3 below), check whether specs/deps.yaml exists and contains ≥ 1 entry; when both conditions hold, invoke /dev-process-toolkit:deps-research <topic> (forked, <topic> = FR title + 1-line summary). The forked skill returns a banner-fenced block of dependency hits drawn from the curated specs/deps.yaml manifest; inject it into context to seed the ## Technical Design interview alongside the spec-research block. Read-only context; never copy verbatim into the FR body. Vacuous path: specs/deps.yaml absent or empty (zero entries) ⇒ skip the invocation entirely (no fork, no tokens), step 3 proceeds as today. Skipped under --no-tech (no ## Technical Design interview to seed; vacuous regardless of manifest state). Shape-violation fallback: block dropped silently, parent continues without the seed (enrichment, not load-bearing — identical to the manifest-absent path). The closing summary at Step 7 MUST emit exactly one literal token (verbatim, no paraphrase) whenever the deps-research subagent fires: ≥ 1 match returned ⇒ MUST emit deps_research_invoked (literal, backticked); zero matches across N entries ⇒ MUST emit deps_research_no_matches (literal, backticked); shape violation ⇒ MUST emit deps_research_shape_violation (literal, backticked).

3. AC prefix — every AC line in the new file uses the shape - AC-<PREFIX>.<N>: <body>, where <PREFIX> is derived via acPrefix(spec) from adapters/_shared/src/ac_prefix.ts: in tracker mode it's the bound tracker ID (e.g., AC-<TKR>-NN.1); in mode: none it's spec.id.slice(23, 29) — the last 6 chars of the ULID's random portion (e.g., AC-VDTAF4.1). Tracker mode requires the tracker: block to be populated before ACs are written (i.e., bind the ticket first, then author ACs). In mode: none, before writing the file, call scanShortUlidCollision(specsDir, spec) from the same module; it throws ShortUlidCollisionError (NFR-10-shape) if another FR already uses the same short-ULID tail. The same short-ULID doubles as the mode-none filename stem, so the collision scan also guards against filename collisions. Never emit literal AC-<digit>.<N> shape — the requirements template carries the placeholder AC-<tracker-id>.<N>; substitute via acPrefix(spec) before writing every AC line.

4. Draft acceptance gate (run BEFORE Provider.sync below). Present the FR draft to the user and ask Approve and proceed? Reply 'y' to create the tracker issue and write the files, 'n' to cancel, 'edit' to revise.

   Marker-detection (STE-262 AC-STE-262.2 — deterministic, runtime byte-grep). The literal byte-string <dpt:auto-approve>v1</dpt:auto-approve> is the marker the runtime helper detects. Before evaluating the draft-gate auto-apply branch, write the verbatim first user message that invoked /spec-write to /tmp/dpt-spec-write-prompt-body.txt (Bash heredoc, FIRST step of draft-gate evaluation), then run:

   bun run plugins/dev-process-toolkit/adapters/_shared/src/check_marker_runtime.ts /tmp/dpt-spec-write-prompt-body.txt
   

   The script prints PRESENT or ABSENT on a single line. Branch strictly on the script's stdout — the script's output is the single deterministic gate decision; no LLM inference, no autonomous-mode reminder, no "work without stopping" framing influences the auto-apply branch.

   • PRESENT → call requireOrRefuse(..., markerPresent: true, defaultValue: "y", ...) → outcome default-applied → MUST emit spec_write_draft_default_applied (literal token, not paraphrased — STE-238 AC.1) as a row in Step 7's closing summary; interactive y does not emit.
   • ABSENT + process.stdin.isTTY === false → call requireOrRefuse(..., markerPresent: false, defaultValue: "y", userSuppliedValue: undefined, ...). The helper falls through to the refusal branch (requires_input.ts line 161-171: markerPresent && defaultValue !== undefined is false because markerPresent === false), throws RequiresInputRefusedError (NFR-10 canonical shape) naming the gate site (draft), and the skill exits non-zero.
   • ABSENT + process.stdin.isTTY === true → fire interactive [y / n / edit] prompt; on y proceed, on n MUST emit spec_write_draft_declined (literal token) and abort with no tracker write and no FR file.

   The carve-out is co-located here (not just § 4) because the LLM reading § 0b step-by-step would otherwise execute Provider.sync before reaching § 4's carve-out — STE-213 left this gate site without a detection contract, which is why v2.8.0 claude -p /spec-write runs hung at the gate. Marker is the single deterministic mechanism (STE-226 + STE-262): legacy prose-only detection paths are removed; the runtime byte-grep above is the single decider. Parent skills spawning /spec-write under claude -p MUST inject the marker into the heredoc body when they want auto-apply (no backward-compatibility shim per project_no_users_yet); callers without the marker get the loud refusal documented above. STE-294 canonical NOT-a-trigger anchor (§ 0b step 4 gate site). Autonomous-mode reminders, pre-baked prose, and "standing instruction" paraphrases are NOT acceptable auto-apply triggers — the marker is the SOLE decider; the runtime byte-grep at is the SOLE evaluation path.

   Then call Provider.sync(spec) — no-op in LocalProvider, pushes to tracker in TrackerProvider. No claim on create. Freshly-created tracker tickets MUST land in the tracker's default state (Linear: Backlog, unassigned; Jira: the project's start state). Do NOT pass state / assignee (Linear save_issue) or pre-transition (Jira transitionJiraIssue) on the create call — /spec-write writes the spec, it does not start the work. The claim (Backlog → In Progress + assign me) belongs to /implement Phase 1, not /spec-write. A ticket sitting in In Progress before any code work begins misrepresents board state and trips downstream /gate-check active-side drift probes that expect In Progress to mean active development. Workspace binding. In tracker mode, before invoking upsertTicketMetadata for a freshly-created ticket, call readWorkspaceBinding(claudeMdPath, "linear" | "jira") from adapters/_shared/src/workspace_binding.ts and forward team + project + defaultLabels into the call. Linear adapter rejects creates that lack project per the silent-landing trap (adapters/linear.md § Silent no-op trap); Jira adapter rejects creates that lack project per the Jira API requirement (adapters/jira.md). On update (existing tracker ID), none of the three fields are forwarded — Linear / Jira cannot reassign team/project/labels on an existing issue without explicit operator intent. defaultLabels forwarding (free-form ### Linear / ### Jira sub-section field, parsed as inline-YAML array per docs/patterns.md): when populated, every entry is forwarded into the create call (Linear: save_issue.labels; Jira: createJiraIssue.additional_fields.labels); when absent or [], no labels field is set and the tracker's default labelling applies. Empty defaultLabels: [] is byte-identical to a missing key (vacuous; round-trip test coverage required).

   Idempotency hardening on Gateway-Timeout retry. upsertTicketMetadata(null, ...) for a freshly-created tracker-mode ticket guards against the orphan-duplicate failure mode an earlier smoke surfaced: the create call returned a Gateway-Timeout while the server-side write succeeded, the single-shot JQL idempotency probe missed the eventual-consistency window, and the retry created a second stub ticket the chain never references. The retry path now widens the idempotency window — on a network-error response (Gateway-Timeout / 504 / connection reset / equivalent) before falling through to a fresh create, retry the JQL idempotency probe with backoff (canonical schedule 1s + 2s + 4s, three attempts). The existing single-shot probe stays as the fast path (no extra latency on the success path). If any backoff probe finds the original write, no duplicate create is issued and the retrieved ticket id is returned. If all three backoff attempts still miss, fall through to a fresh create (the genuine duplicate-create scenario where the original write actually failed server-side) AND surface a one-line warning row in the Step 7 summary table using the tracker_idempotency_uncertain capability key — the operator needs to manually verify before downstream skills bind to the new id. The same contract applies symmetrically to Linear's save_issue (Gateway-Timeout class of failure mode is shared); see adapters/jira.md § upsert_ticket_metadata and adapters/linear.md § upsert_ticket_metadata for adapter-side detail.

   Milestone attachment (any adapter with project_milestone: true — Linear + Jira). After Provider.sync(spec) returns the freshly-allocated tracker ID, read the FR's milestone: frontmatter and call planFileHeadingToMilestoneName(specs/plan/<milestone>.md) from adapters/_shared/src/attach_project_milestone.ts to derive the canonical milestone name (anchor stripped). Then call attachProjectMilestone(provider, project, canonicalName, ticketId) to bind the ticket — Linear binds the native project milestone, Jira read-merge-writes the milestone-<M-token> label. Failure surfaces as MilestoneAttachmentError (NFR-10 canonical shape) — the FR file is not rolled back since the spec is the source of truth. Capability outcome routing (STE-198): attachProjectMilestone returns { capability, createdName? }. Branch on result.capability — null ⇒ no row; "milestone_create_required" ⇒ helper auto-created via mcp__linear__save_milestone (same auth scope as the prior save_issue, no extra prompt) — surface the row naming result.createdName; "milestone_attach_skipped_adapter_limit" ⇒ adapter Schema M project_milestone: false short-circuited the call. Vacuous on mode: none.

5. Post-write self-checks. Immediately after Provider.sync(spec) returns (so the tracker-id is bound), run both in-band checks against the just-written file before /spec-write returns clean:

   • Call runFrontmatterShapeCheck(projectRoot, frFilePath) from adapters/_shared/src/fr_frontmatter.ts. Throws FRFrontmatterShapeError (NFR-10 canonical shape) when the bimodal-identity logic flags the file — the LLM hand-rolled YAML, mutated state, or the helper has a bug. Refusal halts /spec-write before the file leaves a clean signal to downstream skills.
   • Call scanGuessedTrackerIdLiterals([frFilePath, specs/requirements.md]) from adapters/_shared/src/guessed_tracker_id_scan.ts. Any returned violation is a literal AC-<digit>.<N> placeholder that survived substitution; refuse with NFR-10 canonical shape (use the violation's own message field, which is already canonical), naming file:line:column + the offending token + remedy substitute <tracker-id> via acPrefix(spec) and retry.

   The /gate-check probes identity_mode_conditional and guessed_tracker_id stay as the safety net for paths that bypass /spec-write (manual edits, copy-paste from old templates, downstream toolkit consumers).

6. Never write to specs/requirements.md — that file is slimmed to cross-cutting content only.

1. Assess current state

Check which spec files exist in specs/ and how complete they are:

• Read each file and determine: empty template, partially filled, or complete
• Report status to the user

If specs/ doesn't exist, suggest running /dev-process-toolkit:setup first.

2. Determine scope

If $ARGUMENTS specifies a file (requirements, technical, testing, plan), work on that one. If $ARGUMENTS is "all" or empty, work through all files in precedence order:

requirements.md → technical-spec.md → testing-spec.md → plan.md

This order matters because each spec builds on the previous one.

3. For each spec file

requirements.md (WHAT to build — cross-cutting only)

Scope: cross-cutting only. specs/requirements.md captures concerns that span multiple FRs — auth scheme, observability surface, tenancy model, accessibility posture, etc. Per-FR detail lives in specs/frs/<id>.md exclusively. When /spec-write is invoked on a per-FR feature, route the work straight to specs/frs/ (per § 0b) and do NOT touch specs/requirements.md. The requirements-md-no-placeholder gate probe (gate-check #29) flags any ### FR-N: [Feature Name] heading that survives in requirements.md as drift. STE-227 --no-tech: the flag does not affect this cross-cutting flow; only the per-FR technical-spec.md + testing-spec.md interviews below are skipped (the FR body's ## Technical Design and ## Testing sections are written as placeholder lines per § 0b's preamble; the technical reviewer fills them via /spec-write <FR-id>).

Only fire the per-section flow below when the user is filling in genuinely cross-cutting requirements — typically the first time /spec-write runs on a new project, or when an architectural concern emerges that affects multiple FRs.

Ask the questions in this order, one at a time. Wait for each answer before asking the next — do not bundle them into a single turn even when the user is responsive.

Ask what this project is, who it is for, and what problem it solves. Wait for the answer. Then ask what cross-cutting functional requirements exist (auth, observability, accessibility, tenancy — anything that spans multiple FRs). Wait for the answer. Then ask what is explicitly out of scope (project-wide, not per-FR). Wait for the answer. Then ask whether there are any non-functional requirements (performance, security, accessibility).

Per-feature acceptance criteria are NOT collected here — those go in specs/frs/<id>.md via § 0b's FR creation path. If the user names a specific feature, branch into the FR-creation flow rather than expanding § 2 in requirements.md.

Write the answers into the spec using the template structure with per-AC prefixes derived via acPrefix() (tracker ID in tracker mode, short-ULID tail in mode: none). AC lines take the shape AC-<PREFIX>.<N>: ....

Stable anchor IDs: Every ### <PREFIX>: heading you generate or edit must carry its {#<PREFIX>} anchor on the same line, matching the template form ### <tracker-id>: User login {#<tracker-id>} (tracker mode) or ### VDTAF4: User login {#VDTAF4} (mode: none). Same rule applies in plan.md for ## M{N}: ... headings — the {#M{N}} anchor must be present. These anchors are the pointer targets for archival and for cross-references in the traceability matrix, so they must survive heading renames. If you encounter any milestone or FR heading without its anchor, flag it as a warning in the report (step 7) and offer to add it — never silently edit around it.

technical-spec.md (HOW to build it)

Read requirements.md first to understand what needs building. Then ask the questions below in order, one at a time — wait for each answer before asking the next.

Ask what the high-level architecture looks like (read existing code if any). Wait for the answer. Then ask what the key design decisions are and their rationale. Wait for the answer. Then ask what the data model is (schemas, types, database tables). Wait for the answer. Then ask what APIs or interfaces are needed. Wait for the answer. Then ask what the key patterns are (state management, error handling, etc.).

Pre-fill what you can from the codebase and CLAUDE.md. Ask the user to confirm or correct.

testing-spec.md (HOW to test it)

Read requirements.md and technical-spec.md. Then pre-fill the test framework, mocking approach, and file conventions from CLAUDE.md, and identify what NOT to test (generated code, third-party internals).

For the two items that need user input, ask them in order, one at a time — wait for each answer before asking the next. Ask about coverage targets per layer. Wait for the answer. Then ask about the test data strategy (factories, fixtures, seeds, frozen times).

Most of this can be inferred — present your best guess and let the user correct.

plan.md (WHEN to build it)

Read all other specs. Then work through the steps below in order, one at a time — if any step surfaces a question for the user, ask it, wait for the answer, and only then move to the next step. Do not bundle the step-questions into a single turn.

First, break the requirements into milestones (each independently gatable). Then order the milestones by dependency. Then, for each milestone, list tasks in dependency order, acceptance criteria, and gate commands. Finally, draw the milestone dependency graph.

Milestone-number allocation guard. Before claiming any new M<N>, call nextFreeMilestoneNumber(specsDir, changelogPath) from adapters/_shared/src/next_free_milestone_number.ts. The helper does the three-way scan (active specs/plan/, archived specs/plan/archive/, CHANGELOG M<N> references) and returns { next, sources }. Use next as the canonical default. If the user explicitly typed an M<N> argument that appears in any of the three source sets, refuse with NFR-10 canonical shape showing all three breakdowns and the proposed next free number. Never trust a partial ls output and never trust LLM memory: the helper is the single source of truth.

Task Sizing: generated tasks must follow the Task Sizing guidance in templates/spec-templates/plan.md.template — each task ≈ one commit's worth of work, written as a 2-line entry (action line + indented verify: line). If you can't name a single verification step, split the task. See the template's anti-pattern callout for examples of tasks that are too large.

Present the plan and ask for approval.

4. Review and confirm

After completing each spec file: show the user what was written, ask for approval before saving, note any open questions or decisions that need human input.

Draft + commit acceptance gates — marker-driven default-apply (STE-213, STE-220, STE-226). Two gates live in this skill: the FR-draft Approve and proceed? gate at § 0b step 4 (gates tracker create + FR file write), and the commit Apply commit? gate at § 7a (gates the spec commit). Both default-apply y when the prompt body contains the literal line <dpt:auto-approve>v1</dpt:auto-approve> (byte-grep, no inference) — MUST emit spec_write_draft_default_applied and MUST emit spec_write_commit_default_applied respectively (literal tokens, not paraphrased — STE-238 AC.1; the two rows are co-located in the static map at § 7). Interactive n decline at either gate MUST emit the corresponding _declined row (literal token). The marker contract (paste, not paraphrase) is byte-repeated at all three gate sites — § 0b step 4, here, and § 7a — so the LLM reading any single site has the full contract in view. STE-226 replaces STE-213 + STE-220's prose-only detection (legacy Auto Mode Active system-reminder match + claude -p non-interactive inference, both falsified end-to-end across four smoke runs) with this byte-checkable marker; legacy detection is removed, the marker is the single deterministic mechanism, and callers without the marker get interactive gating (no backward-compat shim per project_no_users_yet). Universal pre-commit branch gate (STE-228): § 7a additionally calls requireCommittableBranch({...}) from adapters/_shared/src/require_committable_branch.ts with the name from branchNameFor(...) at skills/spec-write/branch_name_for.ts (new-FR delegates to buildBranchProposal; cross-cutting-only → docs/specs-cross-cutting); outcomes created/edited → git checkout -b; declined → git reset HEAD <paths> rollback + non-zero exit; no-op for off-trunk OR commitType ∈ TRUNK_OK_TYPES = ["ci"]; auto-mode default-apply via the same marker; capability rows branch_gate_* per STE-228 § Branch-name canonical table. Branch-gate literal tokens for every outcome path (STE-238 AC.2). Each outcome MUST emit the matching literal token (backticked) into the Step 7 closing summary:

• gate creates branch interactively ⇒ MUST emit branch_gate_created.
• edited proposal accepted ⇒ MUST emit branch_gate_edited.
• declined ⇒ MUST emit branch_gate_declined.
• marker-driven auto-apply with branch creation ⇒ MUST emit branch_gate_default_applied.
• gate short-circuits because operator was already off-trunk ⇒ MUST emit branch_gate_skipped_already_non_main (NEW token added to the static map at § 7 under STE-238 AC.6).
• remote-probe failure ⇒ MUST emit branch_gate_remote_probe_skipped. The literal token is the byte-checkable signal /gate-check's closing_summary_capability_keys probe greps for; narrative prose like "the gate auto-created the branch" is insufficient.

5. Cross-check consistency

After saving any spec file, automatically check all other existing specs for consistency. Each spec builds on the ones before it, so changes can ripple.

What to check

• requirements.md changed: Check that technical-spec.md covers all functional requirements (architecture, data model, APIs). Check that testing-spec.md has test strategies for all ACs. Check that plan.md milestones cover all requirements and no milestone references removed/renamed FRs.
• technical-spec.md changed: Check that implementation details are consistent with requirements.md scope (no undocumented features, no missing requirements). Check that testing-spec.md mocking and test strategies match the chosen architecture. Check that plan.md task breakdowns match the technical approach.
• testing-spec.md changed: Check that test coverage targets and strategies align with requirements.md ACs and technical-spec.md module boundaries.
• plan.md changed: Check that milestones reference valid FRs/ACs from requirements.md and that task descriptions match technical-spec.md.

How to report

For each inconsistency found, report:

Spec file	Issue	Suggestion
technical-spec.md	<tracker-id> has no architecture section	Add data model for user preferences
plan.md	M2 references removed AC-<tracker-id>.1	Remove or replace with AC-<tracker-id>.2

• If no inconsistencies found, report "All specs are consistent" and move on.
• If inconsistencies found, offer to fix them right now by updating the affected specs. Walk the user through each change and get approval before saving, just like step 4.

6. Risk scan

Before handing off to implementation, do a structured risk scan. Read all specs and the existing codebase, then assess risks using this table:

Category	Risk	Severity	Mitigation
External dependencies		high / medium / low
Breaking changes		high / medium / low
Security surface		high / medium / low
Performance impact		high / medium / low
Data migration		high / medium / low
Unclear acceptance criteria		high / medium / low

Severity levels:

• high — Could block release or cause data loss/security breach
• medium — Significant effort to fix if discovered late
• low — Minor inconvenience, easy to address

Any high-severity risk must be resolved or explicitly accepted before proceeding to implementation.

For each risk found, add it to the relevant spec:

• Technical risks → specs/technical-spec.md (risks/considerations section)
• Unclear ACs → specs/requirements.md (flag the specific AC with a note)
• Testing risks → specs/testing-spec.md (note what's hard to test and the strategy)

If no significant risks found, report "No major risks identified" and move on. Don't invent risks — only flag things that would genuinely surprise someone during implementation.

7a. Stage spec changes and prompt for commit

After Steps 0–6 settle and before Step 7 emits the closing summary, /spec-write stages every file it wrote under specs/ and produces one commit per /spec-write invocation — STE-179 closes the gap that widened setup-bootstrap-committed to toolkit-bootstrap-committed. Subjects (Conventional Commits): new-FR run ⇒ chore(specs): write FR <tracker-id> (tracker mode) or short-ULID stem (mode: none); cross-cutting-only run (pure requirements/technical-spec/testing-spec/plan/M<N>.md edits, no new FR) ⇒ docs(specs): edit cross-cutting specs; hybrid ⇒ new-FR shape (cross-cutting edits land in the same commit). Procedure: stage explicit path list (never git add -A) → diff preview → prompt Apply commit "<subject>"? [y / n / edit].

Marker-detection (STE-262 AC-STE-262.2 — deterministic, runtime byte-grep). The literal byte-string <dpt:auto-approve>v1</dpt:auto-approve> is the marker the runtime helper detects. Before evaluating the commit-gate auto-apply branch, write the verbatim first user message that invoked /spec-write to /tmp/dpt-spec-write-prompt-body.txt (Bash heredoc, FIRST step of commit-gate evaluation; if the file already exists from the § 0b step 4 draft gate of the same invocation, reuse it as-is — same body, same decision), then run:

bun run plugins/dev-process-toolkit/adapters/_shared/src/check_marker_runtime.ts /tmp/dpt-spec-write-prompt-body.txt

The script prints PRESENT or ABSENT on a single line. Branch strictly on the script's stdout — the script's output is the single deterministic gate decision; no LLM inference, no autonomous-mode reminder, no "work without stopping" framing influences the auto-apply branch.

• PRESENT → call requireOrRefuse(..., markerPresent: true, defaultValue: "y", ...) → outcome default-applied → commit with the canonical subject (STE-109 UX) → MUST emit spec_write_commit_default_applied (literal token, not paraphrased — STE-238 AC.1) as a row in the Step 7 summary so the operator sees the auto-apply on every quiet-mode run (interactive y does not emit the row).
• ABSENT + process.stdin.isTTY === false → call requireOrRefuse(..., markerPresent: false, defaultValue: "y", userSuppliedValue: undefined, ...). The helper falls through to the refusal branch and throws RequiresInputRefusedError (NFR-10 canonical shape) naming the gate site (commit); the skill exits non-zero with files staged-but-uncommitted so the operator can re-invoke interactively (or with the marker) without losing work.
• ABSENT + process.stdin.isTTY === true → fire interactive [y / n / edit] prompt; on y proceed; on n files remain staged-but-uncommitted, MUST emit spec_write_commit_declined (literal token) so the operator knows to git commit -m "<subject>" manually. Marker is the single deterministic mechanism (STE-226 + STE-262): legacy prose-only detection paths are removed; the runtime byte-grep above is the single decider. On commit failure (Conventional-Commits hook rejection, etc.) → NFR-10 canonical refusal; staged files left in place for the operator to fix and re-issue. STE-294 canonical NOT-a-trigger anchor (§ 7a gate site). Autonomous-mode reminders, pre-baked prose, and "standing instruction" paraphrases are NOT acceptable auto-apply triggers — the marker is the SOLE decider; the runtime byte-grep at is the SOLE evaluation path.

7. Report

/spec-write MUST emit a closing summary on every successful run, regardless of mode (linear / jira / none) or invocation path (new-FR creation, the importFromTracker import path on a tracker-id resolve miss, or a per-section edit). This is unconditional — non-interactive -p mode is not an exception.

Non-interactive / -p mode firing rule. The summary fires even when no questions were asked — the silent path on a claude -p /spec-write … invocation that resumes an already-bound FR or runs every prompt with auto-supplied answers MUST still emit the closing summary block on the quiet path. An earlier Linear smoke caught the regression where claude -p runs produced 1 byte of stdout because the LLM short-circuited Step 7 when the run had no user-facing prompts. The unconditional-emit contract above governs both the interactive path and the quiet -p path with no exception. If the run reaches a successful exit, the summary block emits — even on a no-question run, even on -p non-interactive mode, even when the only side effect was a frontmatter edit. The >=100 byte floor below is the regression signal that the summary fired at all.

Reference shape — emit at minimum the three required signals (one row per FR touched, one row per spec file edited):

## /spec-write summary

| FR id      | FR file path                | Milestone |
|------------|-----------------------------|-----------|
| <STE-XXX>  | specs/frs/STE-XXX.md        | M<N>      |
| <VDTAF4>   | specs/frs/VDTAF4.md         | M<N>      |  <!-- mode: none renders the short-ULID -->

| Spec file              | Change          |
|------------------------|-----------------|
| specs/plan/M<N>.md     | row added/updated |
| specs/technical-spec.md| edge case backfilled |

Open questions / risks / inconsistencies (if any):
- Which specs are done vs. still need work
- Inconsistencies resolved or still pending
- Risks identified (severity)
- Open questions flagged during the process

Next: Run `/dev-process-toolkit:implement <tracker-id>` when specs are ready.   <!-- new-FR run: recommend the FR-id form (most common, single-FR ship). -->
Next: Run `/dev-process-toolkit:implement <milestone>` when specs are ready.   <!-- cross-cutting-only run (no new FR): recommend the M<N> form (milestone close). -->

Next-line variant rule. When the run created a single new FR, recommend the FR-id form (Run /dev-process-toolkit:implement <tracker-id>) — that's the canonical "ship one FR" path per skills/implement/SKILL.md § Invocation forms (STE-181). When the run only edited cross-cutting specs (no new FR file written), recommend the M form (Run /dev-process-toolkit:implement M<N>) — the operator is presumably finishing a milestone. Hybrid runs (new FR + cross-cutting edit) follow the new-FR shape.

Capability-gap rendering. The "Open questions / risks / inconsistencies" block must render every capability gap as plain prose, drawn from the static plain-language map below — never as a literal AC-<tracker-id>.<N> reference into this toolkit's own internal spec set. The toolkit's AC IDs are opaque jargon to project owners running /spec-write on their own repo (a 2026-04-28 smoke caught the regression: a toolkit-internal AC identifier for the milestone-attach capability surfaced as the rendered description of the gap, replacing what should have been plain prose). Echoing such an identifier inside this section's instructions is itself a regression risk — the LLM may copy it back into the rendered summary; describe failure modes by capability name only.

Static plain-language map (capability key ⇒ rendered prose):

Capability key	Rendered prose
milestone_attach_skipped_adapter_limit	tracker adapter does not support project_milestone — milestone-attach skipped
milestone_create_required	tracker project had zero milestones; created <milestone-name> and attached the FR
workspace_binding_deferred	tracker workspace binding deferred — <MCP> unregistered at /setup time; run /setup --resume-tracker-binding after authenticating
tracker_skipped	tracker step skipped (--code-only flag or pre-flight probe failed) — re-run without flag once MCP loadable to sync
requires_input_refused	step 7b refused — invocation lacked --tracker=<mode> pre-bake; remedy: re-invoke with the flag, or run interactively
plan_only_archival	milestone archived plan-only — no FRs in scope
docs_default_applied	## Docs autonomous default applied — <flag-list> (no pre-baked answers; re-run /setup --migrate to change)
tracker_sync_failed	tracker sync failed — local edits saved, push deferred (re-run /spec-write to retry)
push_ac_unsupported	tracker adapter does not support push_ac_toggle — gate-check will skip the push step
import_acs_empty	imported ticket had zero ACs — TODO marker added to the new FR
workspace_binding_missing	tracker workspace binding absent — ticket landed without team/project association
tracker_idempotency_uncertain	idempotency probe still ambiguous after backoff retry — possible duplicate ticket; operator should manually verify before downstream skills bind to the new id
filename_policy_override	(a) no user override: FR filename derived from tracker policy (Provider.filenameFor) → <filename> (no user override) (b) user override: FR filename derived from tracker policy (Provider.filenameFor) → <filename> (overrode user-proposed: <user-name>)
simplify_tree_dirty	tree dirty after /simplify — modified files: <list>; run /pr or git commit to land the simplification
spec_write_draft_default_applied	/spec-write FR-draft auto-approved (marker dpt:auto-approvev1</dpt:auto-approve> present in prompt body) — verify the draft via specs/frs/<id>.md before /implement
spec_write_draft_declined	/spec-write FR-draft acceptance declined — files not written; re-invoke to retry
spec_write_commit_default_applied	/spec-write commit auto-approved (marker dpt:auto-approvev1</dpt:auto-approve> present in prompt body) — verify diff via git show before /implement
spec_write_commit_declined	/spec-write commit declined — files remain staged, run git commit -m "<subject>" to finish manually
fr_needs_technical_review / fr_technical_review_cleared	flagged: FR <id> needs_technical_review — technical + testing sections are placeholders; run /spec-write <id> (no flag) to complete before /implement. cleared: FR <id> needs_technical_review flag cleared — technical + testing sections completed.
needs_technical_review_label_unsupported / implement_refused_needs_technical_review	unsupported: tracker adapter does not support label push — needs-technical-review label not applied; flag visible only via FR frontmatter. refused: /implement refused — <list of FR ids> flagged needs_technical_review. Run /spec-write <id> (no flag) for each, then re-invoke.
branch_gate_* (created / edited / declined / default_applied / remote_probe_skipped / skipped_already_non_main)	created: gate created branch <branchName> — commit will land there. edited: gate accepted operator-supplied branch name <branchName>. declined: gate declined — skill exited non-zero with zero side effects (no commit, no staging mutation). default_applied: gate auto-created branch <branchName> (auto-mode / -p) — verify before push. remote_probe_skipped: gate could not probe remote branches (offline / no remote / fetch timeout) — local-only collision check used; verify name is unique on remote before push. skipped_already_non_main: gate skipped — operator already off-trunk on <branch>; commit landing here per branch-isolation preference (no new branch created) (STE-238 AC.6)
report_issue_default_applied / report_issue_declined / report_issue_redacted_payload	default_applied: /report-issue gist auto-pushed (marker present in prompt body) — verify the gist contents before sharing. declined: /report-issue gist declined — temp directory deleted, no upload. redacted_payload: /report-issue scrubbed <N> secret-pattern match(es) before upload — see metadata.json for breakdown
spec_research_invoked / spec_research_no_matches / spec_research_shape_violation / deps_research_invoked / deps_research_no_matches / deps_research_shape_violation	spec_research invoked: /spec-research returned <N> related FRs from <M> scanned (active + archived) — context block injected into parent skill. spec_research no_matches: /spec-research scanned <M> FRs (active + archived); zero topic matches — empty block injected into parent skill. spec_research shape_violation: /spec-research output failed shape check (sections / line cap / banner) — block dropped, parent skill proceeds without seed. deps_research invoked: /deps-research returned <N> dependency match(es) from <M> manifest entries — context block injected into parent skill. deps_research no_matches: /deps-research scanned <M> manifest entries; zero matches — empty block injected into parent skill. deps_research shape_violation: /deps-research output failed shape check (banner / sections / line cap) — block dropped silently, parent skill proceeds without seed
tracker_local_reconciled / tracker_local_orphan_local / milestone_local_orphan (STE-284)	reconciled: tracker → local reconciliation imported N FR(s) and M milestone(s) — see specs/frs/<id>.md and specs/plan/<M>.md for the imported content. orphan_local: local FR <id> has no tracker binding — run /spec-write <id> --push-to-tracker to sync, or remove the local file if abandoned. milestone_orphan: local milestone M<N> has no tracker milestone — auto-create via Provider.attachProjectMilestone(M<N>), or remove if abandoned. Auto-import path (preamble reconcile) MUST emit tracker_local_reconciled whenever ≥ 1 FR or milestone is imported; orphan paths MUST emit tracker_local_orphan_local / milestone_local_orphan per offending entity. The byte-checkable tokens are the structural signals /gate-check's probe #49 (tracker_local_reconciliation_drift) and closing_summary_capability_keys cross-reference; narrative paraphrase is insufficient.
tdd_spec_audit_passed / tdd_spec_audit_missing_recovered / tdd_spec_audit_halted	passed: /tdd spec-review audit clean — every AC traced to its implementation and test (drift_count + partial_acs advisory). missing_recovered: /tdd spec-review audit found missing AC(s) on first pass; bounded retry round (test-writer + implementer scoped to missing ACs) recovered them. halted: /tdd spec-review found missing AC(s) and bounded retry did not recover — halt with mode spec-gap; unresolved AC list surfaced. The /tdd orchestrator (or /implement parent) MUST emit exactly one literal token (backticked) in its closing summary — /gate-check's closing_summary_capability_keys probe greps the literal; paraphrase is insufficient.
tracker_config_write_succeeded / tracker_config_write_cancelled / tracker_config_unchanged / tracker_config_write_skipped_adapter_limit / tracker_config_write_mcp_unavailable	succeeded: /setup tracker-config write approved — specs/tracker-config.yaml composed from the active adapter's verbatim status list and the canonical four-role mapping. cancelled: /setup tracker-config write cancelled by operator — no write fired, baseline untouched. unchanged: /setup tracker-config baseline already matches the fresh proposal — no prompt, no write (idempotent re-entry). skipped_adapter_limit: /setup tracker-config write skipped — active adapter does not support list_project_statuses; remedy: switch adapters or author specs/tracker-config.yaml by hand. mcp_unavailable: /setup tracker-config write refused — MCP status fetch threw (server unreachable / auth failure); remedy: re-authenticate the tracker MCP and re-run /setup --resume-tracker-binding (no partial write fires). /setup Step 7f MUST emit exactly one literal token (backticked) per run that reaches this step.
tracker_status_forced	tracker_status_skipped

Annotation: scope = filename_policy_override only. Render variant (a) when the resolver-entry context contains no user-proposed filename (the common pre-baked-stub path — no filename hint in the user's prompt); render variant (b) when the user explicitly proposed an alternative (e.g., the prompt typed specs/frs/foo.md). The row fires on both variants — only the prose differs. mode: none is exempt entirely (no policy-override surface; see the Filename-policy override row paragraph below). <filename> substitutes the actual Provider.filenameFor(spec) output (e.g., STE-179.md for Linear, DST-6.md for Jira); <user-name> substitutes the user-proposed filename verbatim.

Add new keys to this map when a new capability gap surfaces; do not invent ad-hoc prose at runtime, and do not substitute the toolkit's AC-<tracker-id>.<N> ID for the capability key. The map is the single source of truth for capability-gap rendering — bullet bodies are byte-identical across runs. Toolkit-meta vs. user-authored AC IDs. The scrub rule above applies only to this toolkit's own internal AC identifiers (AC-<tracker-id>.<N> references the skill code itself emits about the toolkit's own spec set). User-authored AC references in the active project's FR markdown bodies — legitimate AC-<bound-tracker-id>.<N> entries written by the project owner during the session — pass through unchanged: they are the user's content, not toolkit-meta jargon. If /spec-write is editing an FR file and the user's prose cites a downstream-project AC like AC-XYZ-200.3 as a cross-reference, that reference is preserved verbatim in the rendered summary. The distinguishing test: toolkit-meta IDs are emitted by the skill code; user-authored IDs originate in FR bodies.

Size floor. The summary must be >=100 bytes on stdout — the smoke-test driver guards this via wc -c on the captured log. The two-table-plus-prose shape above clears that floor naturally; do not collapse to a single line. The byte floor is the regression signal that Step 7 fired at all (a prior version of the prose said "Summarize what was completed" and -p mode silently skipped the summary, leaving stdout at 1 byte).

Filename-policy override row. In tracker mode, every successful run that creates a new FR or imports one via importFromTracker(...) MUST surface a filename_policy_override row in the closing summary's open-questions block — regardless of whether the user proposed an alternative filename. The row exists so the operator sees, on every run, that Provider.filenameFor(spec) (e.g., <TKR>-NN.md for Linear, DST-NN.md for Jira) is the authoritative filename source rather than any user-facing name they typed in conversation. An earlier Jira smoke caught the regression where the override was only mentioned when the user's prompt happened to carry an explicit alternative filename; this rule fires the row on every tracker-mode FR write so the signal is unconditional. mode: none is exempt — the short-ULID stem is local-mint via Provider.mintId(), never policy-overridden, so the row is absent on local-mint runs (it would be misleading; there is no override to surface).

Import-path coverage. When importFromTracker(...) ran (resolver step 0a tracker-id/url + findFRByTrackerRef miss), the imported FR appears in the summary table just like a freshly-created one — the operator must see which tracker ID landed in specs/frs/ without filesystem inspection. The summary is the per-skill console-status contract that /setup, /implement, /gate-check, /spec-review, and /simplify all honor; /spec-write was the outlier until the closing-summary contract was added.

Rules

• Ask one clarifying question per turn. Wait for the answer before asking the next. This rule holds at phase transitions too — when two questions look independent, still ask the first, wait, then ask the second. See docs/patterns.md § Pattern 26: Socratic Prompting {#pattern-socratic-prompting} for the canonical rule.
• Socratic Loop Contract (STE-237). Every clarifying Q in this skill — § 1–§ 6 requirement / AC / technical / testing interview prompts — MUST be emitted as an AskUserQuestion tool call (closed-form options OR open-ended with the always-on "Other" free-form fallback), regardless of the autonomous-mode reminder, the auto-approve marker, or pre-baked <command-args> prose. Bare-prose Qs are forbidden. The first-turn contract additionally forbids Write / Edit / NotebookEdit tool calls before the first AskUserQuestion tool_use OR RequiresInputRefusedError raise; Read / Grep / Glob / Bash-read-only orientation is allowed. See docs/auto-mode-protocol.md § Socratic Loop Contract for the full contract.
• Work through specs in precedence order (requirements → technical → testing → plan); each later spec should reference and build on earlier ones
• Ask the user for domain knowledge (don't invent requirements); pre-fill technical details from the codebase and CLAUDE.md where possible
• Present drafts for approval before saving — specs are the source of truth
• Keep acceptance criteria binary (pass/fail, not "good enough")
• Do NOT narrate a specific unallocated tracker ID (e.g., <TKR>-NN) in conversation when drafting — use the literal placeholder <tracker-id> (or the adapter rendering: STE-<N> for Linear, PROJ-<N> for Jira) until the tracker allocator returns the real ID. The placeholder rule for draft files is documented in /spec-write § 0b "Draft with placeholder"; this rule covers the conversational hazard that file-level probes cannot catch.
• NOT-a-trigger anchor (every marker-gated decision). Autonomous-mode reminders, pre-baked prose, and "standing instruction" paraphrases are NOT acceptable auto-apply triggers — the marker is the SOLE decider; the runtime byte-grep at is the SOLE evaluation path. The full forbidden set for the § 0b draft gate, the § 7a branch gate, and every other marker-gated decision in this skill: <system-reminder> text containing "work without stopping", "autonomous-mode" reminders, "standing instruction" paraphrases, pre-baked <command-args> prose, and claude -p non-interactive stdin inference. The runtime byte-grep at adapters/_shared/src/check_marker_runtime.ts is the SOLE byte-checkable evaluation path; the literal marker <dpt:auto-approve>v1</dpt:auto-approve> is the SOLE auto-apply trigger. Anything else is NOT acceptable regardless of how the prompt body paraphrases it. The single-call consolidated arbiter at adapters/_shared/src/gate_marker_refusal.ts (evaluateGateMarkerRefusal({promptBody, isTty, gateSite})) is the canonical decision matrix the § 0b step 4 + § 7a two-step pattern (check_marker_runtime + requireOrRefuse) delegates to in byte-identical form — same primitives, same RequiresInputRefusedError outcome.