Menu
Use when shaping vague, complex, or ambiguous ship-flow requests into a Shape Up pitch, including `/shape`, discussion, or skill-authoring work.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
Use when writing a ship-flow plan for a shaped entity, especially implementation tasks, TDD needs, runtime commands, or scope anchoring. Layer A delegation: superpowers:writing-plans owns plan authoring; ship-flow:test-driven-development owns fallback TDD contracts.
Use when shape finds UI, domain, contract, interface, visual ambiguity, affects_ui, design_required, or no design reference before plan.
Use when ship-flow needs execute-stage implementation from an approved plan, including wave tasks, blocked work, or PR feedback re-entry. Layer A delegation: superpowers:subagent-driven-development owns wave dispatch discipline.
Use when `/ship` should run ship-flow for an entity id or good-enough raw requirement; vague inputs need shape clarification.
Use when verify passed and ship-flow needs review, PR readiness, PR body drafting, or canonical docs sync. Layer A delegation: pr-review-toolkit:review-pr owns PR review persona philosophy.
Use when verifying execute output before ship, including `/verify`, `/ship`, live worktree checks, UI DCs, e2e, reviewer panel, or NIT fixes. Layer A delegation: e2e-pipeline and ship-flow:ui-verify own UI DC verification; reviewer personas own haiku review.
| name | ship-shape |
| description | Use when shaping vague, complex, or ambiguous ship-flow requests into a Shape Up pitch, including `/shape`, discussion, or skill-authoring work. |
| user-invocable | true |
| argument-hint | [--discuss] [directive-text | todo-tid | entity-id] |
You run SHAPE. Output: docs/<wf>/<id>-<slug>/shape.md. Captain has ONE gate per run: confirm / refine / reject.
Shape Up vocabulary (load-bearing โ entity-body schema depends on these names):
problem, appetite, shape_mode, children[], rabbit_holes[], deleted_from_shape[], stated_assumptions[], dag_mermaid.small-batch 2-3d / medium-batch 1-2w / big-batch 6w). Scope fits budget; budget does not flex.docs/<wf>/todos/. Rejected alternative โ claim considered-then-rejected with reason (populates from brainstorming Q-loop or scope-cut during decompose). Shaped child โ vertical E2E slice (pattern: shaped-child). DAG โ mermaid child-dependency diagram; feeds FO Pitch Orchestration.Run before any shape work. Stop and SendMessage(FO) if any check fails.
$WORKFLOW_DIR resolves to a readable docs/ship-flow/ (or equivalent). If unset โ SendMessage(FO): "WORKFLOW_DIR unset โ cannot locate entity folder."planner and executer teammate slots available for this pitch. If TeamCreate fails โ note fallback per Principle 6 Rule A Fallback.answers_density from entity frontmatter. high โ auto-load framework skills per ship-runtime-detect Step R6; skip FO ask. low|vacuum โ SendMessage(FO) with proposed skill list; wait for confirmation./shape "<free text>" โ Mode A (default). /shape --discuss "<text>" โ Mode B Q-loop. /shape <todo-tid> โ promote captured todo. /shape <entity-id> โ shape existing draft.
Escape hatch (skip shape): directive <80 chars AND contains fix | typo | rename | bump | patch | bugfix | hotfix as whole word โ shape unnecessary โ run /ship directly and EXIT. Concrete directive (files / reproducible bug / typed acceptance) โ suggest /ship <requirement> and EXIT.
Skip-when-small (proportionality): even beyond the escape hatch, a single-seam change (one surface, no dependent stages) does NOT need the full FO team pipeline โ it can be version-gate-blocked and is disproportionate. Prefer lean TDD in an isolated worktree + one code-reviewer pass; reserve the full shape โ โฆ โ ship pipeline for multi-surface / dependent-stage work.
Mode auto-routing (on /shape "<text>" without --discuss):
create/build/write/improve/modify a skill, SKILL.md, add a skill; target paths */skills/*, */SKILL.md, .claude/agents/*; entity frontmatter type: skill or domain: skill-authoring.open_questions[] length โฅ 3 โ announce switch and run Mode B.Every proposal records that routing decision as pitch.shape_mode: mode-a | mode-b | mode-c. Absent defaults to mode-a for proposals created before this field existed. Any other value is invalid.
Main agent runs inline. Use TaskCreate to mark phases on main-agent path (skip if a named teammate owns the pitch end-to-end).
Phases: intake โ L0-research โ L1-research โ L2-research โ scope-decompose โ assumption-extract โ appetite-fit-check โ compose-proposal โ cross-review โ captain-gate
| Form | Detection | Action |
|---|---|---|
| Free text | no tid/entity match | use as directive |
| Todo tid | matches docs/<wf>/todos/<tid>.md | read todo body; note tid |
| Entity id | matches docs/<wf>/<id>-<slug>.md OR docs/<wf>/<id>-<slug>/index.md | read entity; use title + body |
Record stage-start ISO timestamp. Resolve WORKFLOW_DIR from docs/*/README.md frontmatter entry-point:. Run escape-hatch check now.
Tracker-issue anchoring (CD5(b) bounded intake-stamping): if the directive references a tracker issue โ a GitHub-style #N reference or a full issue URL (GitHub .../issues/N, Linear issue URL) โ record it in the proposal as pitch.issue (the human-readable ref, e.g. "#49") and pitch.tracker (gh for GitHub, linear for Linear). pitch.issue/pitch.tracker are an all-or-nothing pair: shape-confirm.sh rejects a proposal carrying one without the other, or a pitch.tracker outside gh/linear, before writing anything. shape-confirm.sh then carries both fields forward into the entity frontmatter so the entity is born anchored for plugins/ship-flow/_mods/issue-anchor-guard.md's re-shape detection โ at that later re-shape point, a full GitHub issue URL is accepted end-to-end only when its owner/repo verifies against the local git remote (origin); an unverifiable or genuinely cross-repo URL, or an ambiguous reference, fails visible BLOCK instead. Reference only โ no full tracker integration. Absent a tracker reference, omit both fields; shape-confirm.sh stamps nothing (unchanged behavior).
Run this step immediately after Intake above, and ONLY when Intake matched the Entity id row (an existing docs/<wf>/<id>-<slug> โ i.e. a re-shape via /shape <entity-id> or /shape --discuss <entity-id>). When Intake instead matched Free text or Todo tid โ a brand-new shape with no existing entity yet โ SKIP this section entirely: there is no entity path to resolve, and invoking the resolver here would incorrectly hit its "entity path not found" BLOCK. Gating on Intake's own detection keeps this honest to design premise A1 (issue-anchor-guard is for route-back re-entry on an already-shaped entity, never for new-shape intake).
For a matched Entity id, resolve plugins/ship-flow/_mods/issue-anchor-guard.md for that target entity. If guard emits verdict: narrow or verdict: return, SendMessage(captain) with the source-diff summary before continuing. If no_issue_anchor: true, halt for captain confirmation. See mod for full contract; skip only via explicit --skip-issue-anchor-guard.
Why mandatory: when shape-worker composes shape.md straight from a directive, captain becomes a reviewer who rubber-stamps agent framing โ loses sense of direction, accepts artifacts without owning the problem. Forcing questions reverse the flow: captain articulates verbally first, worker organizes captain's words second. Captain's brain models the answer before speaking; shape.md becomes audit trail of captain's own thinking, not agent inference. AI-augmentation cargo-cult prevention โ do NOT substitute agent thinking for captain thinking. This is codified as INVARIANTS Principle 17: the human review surface is the shape/spec (never the agent-facing plan.md/execute.md); the FO confirms spec content with the captain in plain language rather than offering a plan review.
Small-scope escape hatch (captain agency, NEVER agent's silent judgment): scan /shape <directive> for keywords (case-insensitive, whole-word where applicable): fix typo, rename, bump version, update dependency, remove unused, format, comment (about code comments, not feature comments). If ANY keyword matches, ask captain ONCE:
"This looks like a small-scope change (matched keyword:
{matched}). Skip forcing questions? (Y/n)"
Send via SendMessage(captain) in team mode; inline prompt in bare mode (no team spawned โ degrade per design-officer Bare Mode pattern). Y โ skip articulation, stamp articulation_mode: skipped-small-scope in entity frontmatter, note "Articulation skipped per captain decision (small scope)." in shape.md. n OR no response โ run full forcing flow below.
Forcing questions (default 5) โ send ONE AT A TIME via SendMessage(captain), or inline prompt in bare mode. STOP and wait for each response before asking next. Worker MAY smart-skip a question whose answer is already obvious from the directive, but MUST ask at least 3 of 5:
Worker MUST NOT compose shape.md from agent inference. shape.md is organized from captain's verbatim answers.
Captain Articulation Trail in shape.md โ load-bearing for shape gate review. Captain confirms shape.md is "my words organized", not "agent's draft". Format:
## Captain Articulation Trail
**Q1 (Problem)**: What gets worse without this?
> {captain verbatim answer}
**Q2 (Appetite)**: How long?
> {captain verbatim answer}
(... etc per question asked)
Other shape.md sections (Problem, Scope, Non-goals, Assumptions) MUST reference and quote the Articulation Trail; worker does NOT introduce framing not derived from the trail. The trail is the audit anchor โ if a Layer 1 line cannot be traced back to a captain quote, it does not belong in the proposal.
affected_files, existing_patterns, constraints, prior_entities[โค5], open_questions[].docs/<wf>/_debriefs/*.md (most recent 3-5 files) for recent_warnings[]: issues from ## Issues โ Workflow / ## Issues โ Spacedock, D2-candidates from ## Filed (backlog). Schema: plugins/ship-flow/references/debrief-schema.yaml. Surface any warning directly relevant to the current shape's domain into open_questions[].WebSearch queries only when L0+L1 leave a load-bearing claim unresolved. Usually skip.RULE: L0 via fresh subagent is non-negotiable. Opus 4.7 handles the rest naturally โ don't over-teach.
Appetite is a budget (not estimate): pick small-batch (2-3d, 1-3 children) / medium-batch (1-2w, 3-6 children) / big-batch (6w classic, 5-10 children). Scope fits budget; budget does not stretch. Exceeds big-batch โ flag [EPIC?], recommend sub-pitch.
Vertical-slice children: each child ships E2E standalone. "all-API" / "all-UI" / "every-depends-on-every" = fake decomposition โ re-cut.
Rejected alternatives โ deleted_from_shape[] (field name retained for shape-confirm.sh compat; semantics = "considered but not in scope"). Populate from: (a) brainstorming Q-loop's considered-then-rejected branches, (b) intake clarification (captain said A โ rejected B), (c) scope-cut during decompose (feature trimmed to fit appetite). "Worth doing eventually" โ rabbit_holes[] instead. Musk 5-step procedure intentionally NOT enforced (opus-4.7-naturally-does, MEMORY 2026-04-23); scope protection comes from the appetite-fit check below, not from a forced-delete ritual.
Assumptions: emit stated_assumption per load-bearing claim (schema: plugins/ship-flow/references/entity-body-schema.yaml). Surface any recent_warnings[] from _debriefs/ debrief-read as criticality: moderate assumptions when they affect the current domain. Mandatory: โฅ1 criticality: critical. Run each critical's verification now (30s soft cap); record resolved confidence in confidence_at_shape. Do NOT reproduce verification output in the proposal โ verified_by + bumped confidence_at_shape is the full trace. Raw grep/read results belong to plan-stage research, not shape-stage proposal.
After decomposition and before composing the proposal, verify children fit the declared appetite:
small-batch per child โค 2 days; medium-batch per child โค 3 days; big-batch per child โค 1 week.If fit check fails:
rabbit_holes[], compress child scope, merge thin children. Do NOT extend appetite.[EPIC?] and recommend sub-pitch.Why this exists: opus 4.7 default bias = thoroughness (add) not ruthlessness (cut). Without an explicit fit check, children inflate past appetite and "out of scope" proposals ship. Fit check forces the cut decision BEFORE captain sees the proposal.
Skip for pure bug / UI polish / docs. Run when L0 surfaces drift OR new decision belongs in ARCHITECTURE.md. Draft <!-- section:architecture-impact --> per child; pre-fill before: via bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/extract-map.sh" ARCHITECTURE.md <section>. Uncertain โ emit assumption verified_by: design-contract. Consumer: ship-review planner dispatch patches ARCHITECTURE.md via patch-map.sh --if-hash.
Trigger conditions (any):
## Current Capabilities).Skip when: pitch is pure internal refactor / infra / bug fix with no user-facing surface.
Draft <!-- section:product-impact --> per affected child:
target_section: e.g. capabilities, user-stories, constraints.before / after: current vs new section content.rationale: why this product-level change.Schema: entity-body-schema.yaml โ product-impact block. Consumer: ship-review planner dispatch patches PRODUCT.md via patch-map.sh --if-hash.
Trigger conditions (any):
Draft one <!-- section:readme-impact --> block per affected README section:
target_section: prose heading ("Installation", "Commands", etc.).before / after: current vs new prose.rationale: why.entry_critical: true if the change affects first-time adopter success (triggers cross-review enhanced audit at ship-review).Schema: entity-body-schema.yaml โ readme-impact block. Consumer: ship-review planner dispatch applies via Edit tool + explicit pathspec commit (README is prose-heavy, typically no section tags โ DO NOT use patch-map.sh).
Skip when: pitch is pure internal refactor, bug fix, or infrastructure (no user-facing surface change).
Before composing the final proposal, read the repo-root control-plane docs:
ROADMAP.md โ Now/Next/Later/Shipped context, dependencies, related work,
and follow-up pressure.PRODUCT.md โ product promise, capabilities, user stories, constraints,
"Who It Serves", and "Why It Exists".ARCHITECTURE.md โ only the relevant sections when the pitch touches
schema/API/domain/data-flow/component boundaries/storage/runtime contracts, or
when L0/debrief-read surfaces architecture drift.Record the result in the shape output as canonical intent:
PRODUCT.md and
ARCHITECTURE.md so downstream stages know the omission was deliberate.Do not patch canonical docs during shape. Shape owns intent and captain approval; ship-review owns atomic writes after verify PASS.
Run before composing the proposal:
bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/density-classify.sh" --entity=<path-to-index.md>
Output: high | medium | low | vacuum. Include as answers_density in the proposal JSON pitch block (see schema below). FO uses this to auto-proceed without captain gate when high. Skip when entity index does not yet exist (first-time shape of a directive).
PM skills delegate framing discipline so shape stage doesn't reinvent. Invoke per phase when directive is non-trivial (>80 chars AND not fix/typo/rename/bump/patch/bugfix/hotfix). Skip individual skills with documented rationale; do NOT skip the table wholesale.
| Shape phase | Primary delegate | Supplement (optional) | Feeds proposal field |
|---|---|---|---|
| Intake โ Problem statement | Skill: problem-framing-canvas | Skill: jobs-to-be-done (when directive lacks user-context โ "who hires this for what job") | Problem: |
| Scope decompose โ vertical-slice cuts | Skill: opportunity-solution-tree | Skill: user-story-splitting (9 splitting heuristics โ workflow steps / business rules / data variations / interface options) | Children: + Rejected alternatives: |
| Assumption extract โ criticality filter | Skill: pol-probe-advisor (Point of Leverage โ "which assumption, if wrong, collapses the pitch?") | โ | stated_assumptions[] with criticality: critical |
| Acceptance outcome โ user-observable | Skill: press-release | Skill: user-story (As-a/I-want/So-that structural backup) | Acceptance Outcome: |
Why per-phase delegation, not single pre-compose pass: previously these skills ran once before compose, so criticality: critical filtering happened in LLM head (no POL probe), Children: cuts used opus 4.7 default heuristics (not splitting-heuristics catalog), and Problem: was framed without JTBD's "replaces what" lens. Per-phase delegation forces each framing artifact through a named methodology, not "LLM looked thorough."
Skip rules (per skill, document in proposal ## Shape Report):
Layer A delegate <name> unavailable).--fast flag โ skip all PM framing (escape hatch for tiny pitches).Mandatory (cannot skip): pol-probe-advisor for any pitch with appetite: medium-batch | big-batch. Critical-assumption misfiltering was pitch-103's 18-file/3641-LOC drift root cause; POL probe is the structural guard against criticality: critical decaying into important.
For non-trivial Mode A folder-layout proposals, emit exactly one <!-- section:pm-skill-receipts --> YAML block in shape.md before compose/proposal acceptance. shape-confirm.sh --layout=folder refuses a pitch.shape_mode: mode-a proposal (including an older proposal where the field is absent) that lacks both canonical pitch.pm_skill_receipts and legacy top-level pm_skill_receipts, validates the receipt before any write, then renders the same receipt block into shape.md. Mode B/C are Layer A exceptions: they do not require or render this Mode A receipt block.
The receipt root is pm_skill_receipts with required fields stage: ship-shape, mode: mode-a, appetite: small-batch | medium-batch | big-batch, compose_guard: passed, and one row per required primary delegate: problem-framing-canvas, opportunity-solution-tree, pol-probe-advisor, and press-release. Each primary delegate row must set required: true and the expected phase: problem-framing-canvas โ intake-problem, opportunity-solution-tree โ scope-decompose, pol-probe-advisor โ assumption-extract, press-release โ acceptance-outcome. Optional supplement rows may set required: false; each row records phase, delegate, required, status, evidence, fallback, and rationale.
Allowed status values are exactly:
invoked โ requires non-empty evidence, normally Skill: <delegate>.unavailable โ requires non-empty fallback and rationale.skipped โ requires non-empty rationale; pol-probe-advisor cannot be skipped for medium-batch or big-batch.Hard stop before compose/acceptance: missing receipt block, duplicate receipt blocks, wrong root stage/mode/compose_guard, missing or invalid root appetite, invalid status, missing required delegate row, wrong required delegate phase, primary delegate required: false, missing invoked evidence, missing unavailable fallback/rationale, missing skipped rationale, or medium/big pol-probe-advisor skip. This receipt guard is separate from the 088/097 static Layer A declaration/parity checks; those checks prove skill docs declare delegation, while this artifact proves the shape proposal carried invocation/fallback/skip evidence.
Fat-marker-sketch rule: Children are titles, not specs. Stated assumptions are claims + confidence + criticality + verified_by, not verification RESULTS. DCs, tool choices, file paths, greppable queries, npm-dep choices, LOC estimates all belong to PLAN. If a child description exceeds one line of vertical-slice intent, or if an assumption reproduces verification output, delete detail and keep the claim. Acceptance outcome MUST be user-observable (what captain receives) โ NOT artifact list / infrastructure description / "support for X". One captain-readable claim per pitch.
The captain's gate surface. Structure (present in this order after section-order randomization โ see below):
### Will get โ JTBD-form bullets only. Each bullet MUST use When [situation], [actor] can [verb] [object] form. Each bullet MUST carry a (Check: ...) annotation referencing the paired **W<n>**: line in Layer 2 ### Will-get dogfood checks. No output descriptions ("X surfaces") โ those are not outcomes. C6 enforcement: every **W<n>**: bullet in ### Will get must have a matching **W<n>**: line in Layer 2 ### Will-get dogfood checks or gate hard-blocks.
### Won't get โ explicit no-gos: things a rational captain might assume we'd do but won't. Author judgment, no grep enforcement. At least 1 item on non-trivial pitches.
### Why this scope โ one trade-off line naming the rejected alternative and why. Emit only when appetite/surface divergence is non-trivial. Trivial pitches (escape-hatch threshold) omit this line.
### Captain Bet (mandatory before approve) โ captain-authored substance commitment. Template:
Bet: when this ships, captain expects [observable outcome] within [time window]. If not, this pitch was wrong about [Layer 1 line].
Agent MUST NOT draft and present a filled-in Bet for captain to rubber-stamp. Agent presents the template; captain fills it in. Bet line is always last in Layer 1 (not shuffled). C7 enforcement: Layer 1 MUST NOT contain rubric, 7-factor, or score tokens or gate hard-blocks.
/goal-ready phrasing (optional): when the captain may drive the pipeline under a Claude Code /goal, phrase the Bet's [observable outcome] as a verifiable check (a command + its pass/fail result, e.g. "bash โฆ/test-X.sh exits 0 and PR is merged") and append a turn-bound (or stop after N turns). A Bet phrased this way drops straight into /goal <condition> and stays honest under the transcript-judged evaluator. See INVARIANTS "Evidence discipline under an active /goal".
Section-order randomization: at compose time, shuffle the order of ### Will get, ### Won't get, ### Why this scope using a per-gate seed (e.g., sha256 of entity-id + gate-timestamp, take last 4 hex digits mod 6 for permutation index). ### Captain Bet is always last โ not shuffled. Purpose: defeats agent-pattern-match templating across gates, where agent learns "Will get always first โ approve on structure".
expand <section> request)Not auto-expanded. Captain must explicitly type expand <section> to receive it. Includes: pre-mortem, cross-review rubric score, PM artifacts, ### Will-get dogfood checks (paired W checks), problem evidence, assumptions, rejected alternatives, open questions. Agent presents Layer 1 only at gate; Layer 2 is available on request.
ONE block โ captain's only view until gate:
Pitch proposal: <title>
Problem:
<1-3 sentences โ gap, who feels it, why now. No solution language.>
Acceptance Outcome (what the captain GETS when this ships โ user-observable):
<1-3 sentences. NO internal artifacts. NO "infrastructure for X". Anchor for confirm/refine.>
Appetite: <small-batch | medium-batch | big-batch> (<concrete time budget>)
Children (N, each vertical E2E that ships standalone):
<id>.1 โ <title> (deps: none)
<id>.2 โ <title> (deps: <parent-slug>)
Rabbit holes (auto-captured to docs/<wf>/todos/ on confirm):
- <one-line>
Rejected alternatives (NOT captured โ rationale for the record):
- <claim> โ <reason>
Stated assumptions:
A1 (critical, <conf>%): <claim>
Pre-mortem (1 sentence, โค30 words; pick exactly one category):
<category>: <projected failure mode if pitch ships per spec but doesn't deliver value>
DAG:
```mermaid
graph LR
A[<child-1>] --> B[<child-2>]
Confirm / refine: "" / reject ?
Mermaid fence MUST start with `graph` (shape-confirm.sh requires it).
#### ASCII DAG for chat presentation (captain UX)
When presenting the captain gate proposal IN CHAT, render the DAG as an **ASCII art diagram** (boxes + arrows), NOT mermaid. Mermaid renders only in artifacts (`shape.md`, plan.md), not in the captain's terminal. Captain UX preference (codified 2026-04-29 during pitch-108 shape session): "ASCII for chat, mermaid for files."
Pattern:
โโโโโโโโโโโโโโโโโโโโโโโโ โโโโโโโโโโโโโโโโโโโโโโโโ โ .1 โ โโโโโโโถ โ .2 โ โ โ depends โ โ โ ~d โ on โ ~d โ โ โ โ โ โ โข bullet 1 โ โ โข bullet 1 โ โ โข bullet 2 โ โ โข bullet 2 โ โ โ โ โ โ DC: โ โ DC: โ โโโโโโโโโโโโโโโโโโโโโโโโ โโโโโโโโโโโโโโโโโโโโโโโโ Wave 1 (foundation) Wave 2 (verifier)
For 3+ children with branches, use:
โโโโโโโโโโโโ
โโโถโ <id>.2 โ
โ โโโโโโโโโโโโ
โโโโโโโโโโโโ โ โ .1 โโโโค โโโโโโโโโโโโ โ โ โโโโโโโโโโโโ โโโถโ .3 โ โโโโโโโโโโโโ
The artifact (`shape.md` / plan.md / index.md) ALWAYS uses mermaid โ `shape-confirm.sh` parses mermaid and `dag_mermaid` field in proposal JSON requires mermaid syntax. ASCII is a chat-only render.
### Pre-mortem (mandatory on non-trivial pitch; before cross-review)
After composing the proposal, write **one sentence** answering: "If this pitch ships exactly per spec but post-ship inspection finds it isn't delivering value, what's the single most likely cause?" Append as `pre_mortem` (1 sentence, โค30 words; pick **exactly one** category):
- **wrong-problem** โ acceptance outcome solved, but problem misframed (solving the wrong gap).
- **wrong-dcs** โ children pass verify but their aggregate doesn't deliver the acceptance outcome (vanity DCs).
- **wrong-framing-lens** โ engineering decomposition applied to a relational / aesthetic / macro problem; Musk-lens mismatch (route to Naval / Buffett / Covey / Feynman / Dalio context next time).
- **hidden-dependency** โ load-bearing assumption stayed implicit (never made it to `stated_assumptions[]`).
- **over-conviction** โ proposal text contains "ไธๅฎ / ้กฏ็ถ / ๅฟ
้ / clearly / obviously" without Loss Function check; captain bias signal per MEMORY Conviction Calibration.
**Skip-when**: same threshold as escape-hatch (directive <80 chars AND fix/typo/rename/bump/patch/bugfix/hotfix). Otherwise mandatory.
**Why this exists**: Cross-review uses the same 6-factor rubric the proposer used to compose. Both can miss the same blind spot (composing-rubric overlap). Pre-mortem forces a future-failure perspective the composing rubric cannot generate. The `over-conviction` category is the named hook for MEMORY's "ไธๅฎ / ้กฏ็ถ / ๅฟ
้ " signal โ surfaces captain bias as a structural gate, not a vibe check.
### Cross-review gate (before captain gate) โ Principle 6 Rule C
Dispatch cross-review to `executer` teammate. **Reviewer model fallback when no team**: fresh **sonnet** by default; upgrade to fresh **opus** when `appetite: big-batch` (scope warrants heavier independent review). Apply the 7-factor rubric (per INVARIANTS Principle 6 Rule C #106 T1.3 + pre-mortem extension): **feasibility** appetite-fit within budget / **executable scope** true E2E vertical / **quality** rejected alternatives โฅ1 + critical assumption โฅ1 + appetite-fit check ran / **DC adequacy** observable done-checks / **canonical sync** architecture-impact block when ARCHITECTURE.md affected / **Reverse-audit previous stage** (no prior stage at shape โ ask: does the spec expose any design constraint the prior debrief flagged as a gap?) / **pre-mortem credibility** (does the named category match the most plausible failure mode? does the one-sentence projection survive a "wait, isn't X more likely?" challenge? rote-pick of `hidden-dependency` without supporting reasoning = WARN), rating each PASS/WARN/FAIL, then emit verdict: **PROCEED** โ present to captain; **VETO** โ silently loop to scope-decompose with feedback; **PROMPT_CAPTAIN** โ present proposal + reviewer concern together.
**Proposal budget**: the proposal text passed to the cross-reviewer MUST be โค400 words. Longer = detail creep; trim BEFORE dispatching cross-review, not after.
**Circuit breaker**: if the cross-review teammate is unresponsive (phantom team / SendMessage timeout / fresh-Agent stall), fall back per INVARIANTS Rule A Fallback โ fresh sonnet by default, fresh opus when `appetite: big-batch`. Do not block on an unresponsive reviewer.
---
## Mode B โ Interactive Q-loop (Layer A exception)
**Trigger**: `/shape --discuss "<text>"` OR auto-routed when L0 returns โฅ3 `open_questions`.
**Delegation**: `superpowers:brainstorming` owns the HARD-GATE Q-loop. Do NOT re-teach.
**Flow**: announce mode โ `Skill: superpowers:brainstorming` โ on completion apply Mode A Layer B wrap on the brainstorm report (appetite + vertical-slice decompose + appetite-fit check + โฅ1 critical assumption + DAG + architecture-impact if needed). Rejected-alternatives populate from brainstorming's Q-loop considered-then-dropped branches. โ cross-review gate โ present proposal โ captain gate.
Exception rationale: brainstorming's Q-loop handles discovery; Shape Up framing (appetite/DAG/deletes/assumptions) is Layer B, not in superpowers.
## Mode C โ Skill-authoring (Layer A exception)
**Trigger**: keywords / path patterns / frontmatter (see Mode auto-routing ยง1).
**Delegation**: `superpowers:writing-skills` owns skill design + claude 4.7 knowledge + RED/GREEN/REFACTOR + frontmatter discipline. Do NOT re-teach.
**Flow**: announce mode โ `Skill: superpowers:writing-skills` โ on completion apply Layer B wrap: appetite (`small-batch` single SKILL.md / `medium-batch` multi-file); **children = 1 per SKILL.md (default)**; decompose into โฅ2 children ONLY when (a) design spans **different SKILL.md files** OR (b) supporting **scripts/tests need separate landing** than the skill (e.g., lib script lands before skill references it). Assumptions extracted from writing-skills invariants (frontmatter valid, description matches trigger, etc.), โฅ1 critical. Architecture-impact only if new skill reshapes plugin structure. Cross-review gate โ present proposal โ captain gate.
Exception rationale: skill design + 4.7 knowledge is writing-skills' domain; Shape Up framing is Layer B.
---
## Captain response (all modes)
### Confirm
1. **Allocate the pitch ID** โ ship-flow's native, worktree-aware, reservation-based allocator. Replaces the removed spacedock python `commission/bin/status --next-id` (spacedock v1 makes `--next-id` non-applicable for `id-style: slug`; ship-flow owns its `<N>-<slug>` + `.N`-children numeric scheme):
```bash
bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/allocate-id.sh" "$WORKFLOW_DIR"
Echoes the next top-level number โ pitch ID. Child IDs = <pitch-id>.N (dense, no gaps). The allocator scans all git worktrees + _archive + pending reservations and writes a reservation before returning, so the allocate โ commit gap no longer races (the number is held for up to 2h or until the entity materializes โ the old "one uninterrupted pair" discipline is no longer required).
Serialize โ temp JSON โ invoke atomic writer:
bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/shape-confirm.sh" --proposal="$PROPOSAL_JSON" --layout=folder --workflow-dir="$WORKFLOW_DIR"
--layout=folder (default for new pitches) writes docs/<wf>/<id>-<slug>/index.md + shape.md.
Report: 1 pitch (folder) + N shaped-children + M rabbit-hole todos + ROADMAP.md rows, ONE commit SHA.
Captain Bet capture (mandatory on non-trivial pitches before presenting Layer 1):
Before presenting Layer 1 to captain, verify ### Captain Bet template line is present in the Layer 1 draft. Gate refuses to advance if Bet line still contains unfilled template placeholders ([observable outcome], [time window], [Layer 1 line] verbatim) โ those are signals that agent drafted it and captain hasn't committed.
After captain approves and fills in the Bet line, copy the captain-authored Bet verbatim into the entity body under ## Captain Bet (gate approval <YYYY-MM-DD>) block. Captain MAY edit any agent-drafted Layer 1 bullet inline, but the Bet line MUST be captain-authored โ agent cannot draft a filled-in Bet and then captain rubber-stamp it.
Retro prompt (re-read at ship + 2 weeks): "Did the Bet match outcome? YES / NO / PARTIAL. If NO: which Layer 1 line was wrong?" Track match rate across pitches โ 3 consecutive Bet โ Outcome retros trigger the kill criterion (freeze format rollout, re-shape entity 109's successor).
Skip when: escape-hatch (directive <80 chars AND fix/typo/rename/bump/patch/bugfix/hotfix). Otherwise mandatory.
Refine โ re-run research + decompose with refinement appended (lean re-run; don't diff-patch). Max 2 rounds; then ask: refine / save draft / reject. Reject โ do NOT invoke shape-confirm.sh; verify git status --short clean; emit Pitch rejected. No files written. and EXIT.
On first /shape of a new pitch, spawn team so /ship / /verify reuse hot context. Default: planner (opus) + executer (sonnet) + verifier (opus or sonnet by pitch size). When pitch trigger fires affects_ui:true OR domain: OR design_required:true OR contract_decision_required:true OR file globs *.tsx|*.css|*.html OR explicit --design flag, also spawn designer (opus) โ gates the conditional design stage between shape and plan.
# Default spawn (all pitches) โ every named teammate is a spacedock:ensign unit (canonical worker primitive):
TeamCreate(team_name: "pitch-<id>", members: ["planner", "executer", "verifier"])
Agent(subagent_type: "spacedock:ensign", team_name: "pitch-<id>", name: "planner", model: "opus", task: "Planner for pitch-<id>. Resolve the shape artifact with plugins/ship-flow/lib/resolve-shape-artifact.sh; read canonical shape.md or legacy spec.md fallback.")
Agent(subagent_type: "spacedock:ensign", team_name: "pitch-<id>", name: "executer", model: "sonnet", task: "Executer for pitch-<id>. Atomic commits, DC-first.")
# Conditional spawn โ when design-trigger fires (affects_ui:true OR domain: OR design_required:true OR contract_decision_required:true OR *.tsx|*.css|*.html glob OR --design flag):
Agent(subagent_type: "spacedock:ensign", team_name: "pitch-<id>", name: "designer", model: "opus", task: "Designer for pitch-<id>. Resolve the shape artifact with plugins/ship-flow/lib/resolve-shape-artifact.sh; route Category 0/A/B/C/D and registered domain lanes; emit design.md plus the narrow artifact bundle required by the selected design-dispatch-manifest.")
SendMessage(to: "planner", body: "Proceed to /plan for pitch-<id>. Resolve the shape artifact with plugins/ship-flow/lib/resolve-shape-artifact.sh; output plan.md.")
Conditional design-stage trigger (FO check, runs after shape-confirm.sh):
affects_ui:true field set in entity frontmatter (sharp-stage-set; default false), ORdomain: field set in entity frontmatter (sharp-stage-set via registry-classify; default unset), ORdesign_required:true field set in entity frontmatter for schema/API/domain/architecture contract impact, ORcontract_decision_required:true field set in entity frontmatter because open_contract_decisions[] is non-empty, ORFiles modified or architecture-impact block citing path matching glob *.tsx | *.css | *.html, OR--design flag on /shape invocation.The FO ALWAYS advances the entity to the design stage (skill: ship-flow:ship-design) โ per INVARIANTS Principle 11, design always runs and the former skip-when clause is removed; C14's transition graph has no shapeโplan edge. This trigger selects the design lane, not skip-vs-run: when ANY condition fires, design does full lane work (UI / domain / contract); when NONE fires, the entity walks ship-design's Phase 0 trivial-pass fast-path (minimal design.md + unconditional PROCEED) and then advances to plan. Run ## Phase 8.5 โ Domain Registry Validation before evaluating the trigger so non-UI domains are not misclassified because domain: was never recorded.
Stage continuation โ SendMessage to named teammate (~10ร faster than fresh dispatch). Fresh-subagent reserved for Rule A exceptions: (a) adversarial review across teammates; (b) clearly separate domain; (c) explicit captain request; (d) cross-review gate between stages.
Top-level keys: pitch (with id, slug kebab โค40, title, problem, acceptance_outcome (โฅ50 chars, user-observable, mandatory), appetite, shape_mode (mode-a | mode-b | mode-c; absent defaults to mode-a; unknown values reject before writes), pm_skill_receipts (required for non-trivial Mode A folder-layout proposals; root includes stage, mode, appetite, compose_guard, and primary delegate rows), stated_assumptions[], dag_mermaid โ first line MUST start with graph, pre_mortem (mandatory on non-trivial pitch; object with category enum: wrong-problem | wrong-dcs | wrong-framing-lens | hidden-dependency | over-conviction and one_liner โค30 words)), children[] (id = <pitch.id>.<N> dense no gaps, slug, title, vertical_slice, depends_on[] via child slugs), rabbit_holes[] (slug, claim, domain, guess_files[]), deleted_from_shape[] (claim, reason โ semantically "rejected alternatives"; SHOULD have โฅ1 on non-trivial pitch; empty = captain may have under-shaped, warrants cross-review PROMPT_CAPTAIN). Legacy top-level pm_skill_receipts remains accepted by shape-confirm.sh for compatibility, but new Mode A proposals should write pitch.pm_skill_receipts. Mode B/C proposals omit it. stated_assumptions[] item: id, claim, verified_by (codebase-grep | lib-docs | web-search | design-contract | skill-source-read), verification (bash), confidence_at_shape (0-100), criticality (critical | important | nice-to-know) โ MUST have โฅ1 critical. Full entity-body semantics: plugins/ship-flow/references/entity-body-schema.yaml. answers_density (optional, high | medium | low | vacuum): emit when pre-classified; omit to defer lazy classification.
pre_mortem โ cross-review VETO (loops to scope-decompose). Anchor: composing-rubric blind spot โ proposer + reviewer share the 6-factor rubric, so pre-mortem provides an orthogonal future-failure axis the composing rubric cannot generate. Conviction Calibration tie-in: over-conviction category is the named hook for MEMORY "ไธๅฎ/้กฏ็ถ/ๅฟ
้ " signal.acceptance_outcome OR <50 chars โ silent foundation-only-ceremony risk; shape-confirm.sh rejects with exit 10. Source: pitch 096 demo (3 of 4 children would have shipped placeholder files nothing consumed).shape-confirm.sh only; no direct entity/ROADMAP edits; no -a/-A staging.architecture-impact / product-impact / readme-impact) โ silent drift at ship-review (planner dispatch has nothing to patch).--next-id โ shape-confirm.sh commit = ONE uninterrupted pair (MEMORY #5).git status clean).git add <path> && git commit ... -- <path>.shape-confirm.sh invocation time, git rev-parse --absolute-git-dir MUST resolve under .claude/worktrees/ (not the repo's main .git/). On main tip โ HALT, spawn worktree via Claude Code EnterWorktree (or git worktree add .claude/worktrees/shape-<id>-<slug> -b worktree-shape-<id>-<slug> manually) before running confirm path. Rationale: pathspec-lock is necessary but NOT sufficient under parallel-session contention.After captain confirms the spec, write the ### Hand-off to Design block in the entity body. This is the explicit hand-off consumed by the design stage Boot Self-Check.
Read the incoming hand-off from the prior stage (none for shape โ this is stage 1).
Emit ### Hand-off to Design:
ui_surfaces: list visible UI surfaces inferred from spec (tabs, sidebars, buttons, forms)framework_detected: run Skill: ship-flow:ship-runtime-detect Step R5 โ record framework=X theme_indirection=Y design_canonical_dir=Zopen_design_questions: unresolved design decisions from spec Scope In (e.g., color tokens, layout choice)open_contract_decisions[]: unresolved non-UI contract/interface choices from spec or issue text. Emit this whenever there are 2+ viable semantics and planner must not choose silently. Examples: selector grammar (find role <r> --name "<v>" vs CSS attributes vs {role,name} object), API vocabulary, tool protocol, DSL syntax, schema/message format, mapperโnative boundary contract.pm_framing_output: reference to problem-framing-canvas or press-release artifact path if run in shapeIf open_contract_decisions[] is non-empty, set contract_decision_required: true in entity frontmatter and advance to design even when affects_ui: false and no domain registry exists. The design stage owns the trade-off table and captain decision; plan must not self-select the canonical grammar/protocol/schema/API vocabulary.
If affects_ui: false in entity frontmatter โ omit ui_surfaces and framework_detected; set open_design_questions: [] unless the entity has non-UI contract/interface ambiguity, in which case emit open_contract_decisions[].
Design trivial-pass hand-off (G14 amended by contract-design gate; INVARIANTS Principle 11 โ design always runs): shape ALWAYS emits ### Hand-off to Design and advances to the design stage. When affects_ui: false AND domain: unset AND design_required: false AND contract_decision_required: false, design runs its Phase 0 trivial-pass fast-path โ it is NOT skipped at the pipeline level (no shapeโplan transition; C14 rejects it). If any signal is set, design does full lane work.
When affects_ui: false AND domain: is unset AND design_required: false AND contract_decision_required: false โ shape's ### Hand-off to Design omits ui_surfaces and framework_detected; sets open_design_questions: []; sets open_contract_decisions: []. The ### Hand-off to Plan with design-skipped: true (single field) is then emitted by ship-design's Phase 0 trivial-pass (INVARIANTS Principle 11.3), NOT by shape โ plan Step 1.6 reads it from design.md. Rationale: the design-skipped: true marker lets plan distinguish "design intentionally bypassed (trivial-pass)" from "block missing โ error", while the entity still passes through the uniform shape โ design โ plan chain. Validated by check-invariants.sh --check plan-imported-design-dcs-emitted.
After spec is composed (or post-shape-confirm if pre-existing spec), run or instruct the shape worker to run domain classification once whenever the pitch may touch non-UI domains such as schema, saga, RBAC, permissions, data model, migrations, storage, API contract, workflow runtime, or registry-backed contract domains.
Registry-backed contract domains are no-UI/no-schema design triggers. The portable plugin-default set is agent-contract, api-vocabulary, selector-grammar, tool-protocol, dsl, and message-format; these domains can route a pitch to design from registry evidence even when there is no UI or schema trigger.
Before classification, check whether adopter routing exists:
test -f .claude/ship-flow/domains.yaml
test -f .claude/ship-flow/skill-routing.yaml
If .claude/ship-flow/skill-routing.yaml is missing, run:
bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/discover-adopter-skills.sh" --root=.
Present the discovered routing draft to the captain and record it in
### Project Skills when accepted. Missing routing is not a hard block for
small fixes, but non-trivial multi-surface pitches MUST surface the gap before
plan so planner does not collapse adopter-specific skills into generic
project-db / fmodel defaults.
bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/registry-resolve.sh" --classify <spec-or-entity-file>
Branch on output:
status=ok + non-empty matched=<domain-list> โ matched can be a single domain or a comma-separated list; split the matched list on commas, trim names, set domain: <name> to the first matched domain at shape stage, and advance entity to design stage via standard Hand-off to Design path.status=partial_coverage + matched=<dom-list> + missing=<dom-list> โ set domain: <name> to the first matched domain at shape stage; emit ## Domain Classification Report block listing partial coverage (ship-design router will surface partial-coverage annotation).status=ok + empty matched= โ no domain match; do NOT set domain: (UI-only or no-trigger path proceeds as before).status=parse_error / invalid_trigger_config (M4 exit 20 / M5 exit 21) โ fail loud per INVARIANTS Principle 9; BLOCK shape. Do not proceed to design or plan.When any matched domain is one of the registry-backed contract domains above, also set contract_decision_required:true from registry evidence and emit at least one open_contract_decisions[] item naming the contract/interface choice that design must resolve. This applies even when matched= contains comma-separated multi-domain matches and the first domain: value is another contract domain. This prevents planner from silently choosing selector grammar, API vocabulary, tool protocol, DSL syntax, message format, or agent contract semantics.
If captain/shape explicitly specifies a domain: value before or during shape, validate that value before the shape gate:
bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/registry-resolve.sh" --validate --domain=<domain>
Branch on validation output:
status=ok โ preserve domain: <name> in entity frontmatter and continue to design hand-off.status=specialist_missing (M1), status=knowledge_module_missing (M2), status=parse_error (M4), or status=invalid_trigger_config (M5) โ BLOCK at shape gate with HALT-with-options; do not push ambiguity downstream to design.Shape output/frontmatter evidence MUST include a grep-friendly Domain Registry Validation block:
## Domain Registry Validation
- classify: bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/registry-resolve.sh" --classify <spec-or-entity-file>
- validate: bash "${CLAUDE_PLUGIN_ROOT:-plugins/ship-flow}/lib/registry-resolve.sh" --validate --domain=<domain>
- domain: <name>
- result: proceed | HALT-with-options
For blocked validation, HALT-with-options must name the registry status and offer the same captain-visible choices design uses as second-line defense: skip, generalist-marker, or file-specialist-first. Preserve ship-design HALT behavior; shape-time validation catches the problem earlier, design remains the second line of defense.
If a later rerun or downstream design-stage validation supersedes a stale shape HALT (for example, knowledge_module_missing from an old plugin cache, then current registry validation returns status=ok), do not delete the original evidence. The downstream artifact MUST add a grep-friendly resolution block:
### Registry Validation Resolution
- prior_result: <shape status/result>
- current_result: ok
- resolution: superseded_by_design_stage_validation
- reason: <specialist/knowledge module now present, or adopter config corrected>
Plan stage consumes the latest resolved registry result only when this resolution block is present; otherwise a shape HALT-with-options remains blocking evidence.
Multi-domain match disambiguation v1: pick matched[0] (first registered domain name). v2 multi-domain dispatch is out of 113.1 scope.
Require a ### Metrics subsection in ## Sharp Report / shape-stage report output.
Use grep-friendly key: value lines:
status: passed | rejected | blockedduration_minutes: wall-clock minutes for shape/sharpiteration_count: captain Q-loop or refine iterationspath: shape+sharp | sharp-only | escape-hatchopen_contract_decisions_count: unresolved contract choices handed to designdomain_matches_count: registry domains matched at shape timeplugins/ship-flow/references/entity-body-schema.yaml.plugins/ship-flow/lib/shape-confirm.sh (--layout=folder lands Wave 5 of #085).plugins/ship-flow/skills/add-todos/SKILL.md.plugins/ship-flow/_mods/reverse-recovery-audit.md (plugin-canonical); adopter override docs/ship-flow/_mods/reverse-recovery-audit.md when present.plugins/ship-flow/_mods/issue-anchor-guard.md.superpowers:brainstorming (Mode B), superpowers:writing-skills (Mode C).plugins/ship-flow/INVARIANTS.md (context continuity + 3-layer architecture + cross-review).plugins/ship-flow/references/entity-body-schema.yaml โ stages.sharp.hand_off_to_design.