SLICE BUILDER PROMPT

# Slice Builder Prompt

**Purpose:** Expand a one-line feature description into a complete filled vertical slice — all 11 docs from `docs/slice-templates/` at gold-standard fidelity, ready to hand to an implementation agent.

**Meta-design note:** This prompt is tuned for Opus 4.7. It front-loads the ambiguity-resolution step (because one-liners are woefully underspecified), freezes high-risk contracts early (persona, journey, DDL, API), derives the rest in parallel, then self-validates against the 10 validator agents before declaring done.

---

## How to invoke

Paste everything between the `<<<PROMPT>>>` markers below into a fresh Claude Code session. Replace `{{ONE_LINER}}` with your feature description (1-2 sentences). Optionally set `{{SLICE_ID}}` or let the agent generate one.

Example invocations:

- `{{ONE_LINER}}` = `A CSM can flag a customer as at-risk and get an auto-generated win-back playbook.`
- `{{ONE_LINER}}` = `A finance analyst can close a period and export a GL trial balance with variance explanations.`
- `{{ONE_LINER}}` = `A PM can group related features into an Initiative and see aggregated progress rollup.`

---

<<<PROMPT>>>

# Role

You are a senior product+engineering lead expanding a one-line feature description into a complete production-quality vertical slice for the current codebase.

Your deliverable is **11 filled slice documents** at `docs/slices/{{SLICE_ID}}/` that pass the `slice-validator-orchestrator` agent's checks, plus any supporting artifacts (migrations, seed TypeScript, assertion SQL, Gherkin features).

# Inputs

- **ONE_LINER:** `{{ONE_LINER}}`
- **SLICE_ID:** `{{SLICE_ID or "derive from ONE_LINER, kebab-case, suffix -v1"}}`

# Hard constraints (non-negotiable)

1. **Read the templates first.** Every doc you write is a filled copy of the corresponding template in `docs/slice-templates/*.template.md`. Do not invent structure. Do not skip sections marked `[REQUIRED]`.

2. **Read the validator rules.** Each template has a "Validator Will Fail If …" block and a paired agent at `.claude/agents/slice-validator-<name>.md`. Your output must pass every rule before you declare done.

3. **Banned phrases.** These MUST NOT appear anywhere in any doc: `drives growth`, `seamless`, `powerful`, `best-in-class`, `streamlined`, `cutting-edge`, `robust`, `intuitive`, `game-changing`, `synergy`, `leverage`, `next-gen`, `world-class`, `enterprise-grade`. Rewrite until they're gone.

4. **Specificity floor.** Personas must be named with age, tenure, 5+ tools, literal quotes. Seed data must use recognizable names (Acme, Wayne — not "Customer 1"). JTBDs must have numeric metrics. Every state matrix has all 7 states filled. Every API endpoint has a schema. Every table has `org_id` or `tenant_id`.

5. **Codebase conformance.** Match your project conventions from `CLAUDE.md`. The example conventions the skills ship with assume: ESM imports with `.js` extensions, `{ data }` response envelope, raw `pg` (no ORM), `getRequiredOrgId()` from a shared utils package, snake_case DB, Zod schemas, monorepo services under `apps/services/`, fail-closed gateway routes in `route-permissions.data.ts`. Override these by editing your `CLAUDE.md`.

6. **Reference the example slice.** `docs/slices/00-example-bookmark-v1/` contains filled `01`, `02`, `09` at the required fidelity. Match or exceed that level of detail.

7. **If you cannot make a decision confidently, ASK.** Do not silently pick. See Phase 1 below.

# Execution phases

## Phase 1 — Intake & ambiguity resolution (blocks everything)

The one-liner underspecifies at least these dimensions. Ask ONLY about ones the ONE_LINER genuinely leaves ambiguous; skip anything the one-liner already decides.

Use the `AskUserQuestion` tool. **Maximum 5 questions, single batched message.** If the user won't answer or says "you decide", proceed with documented assumptions (flag them in every doc's frontmatter under `assumptions:`).

Required to lock before Phase 2:

1. **Which persona archetype?** (PM, CPO, CSM, Engineering Lead, Sales Ops, RevOps, Finance, Support, Admin, End User) — the ONE_LINER may imply one; confirm.
2. **Which service owns the new entity/entities?** (one of your monorepo services). Look at the nearest domain in `apps/services/` — quote service name.
3. **Does this slice involve AI?** (If yes: assistive / augmented / autonomous tier per `your AI architecture doc, if any`.) If ONE_LINER doesn't hint, default to `status: n/a` for 08-AI-SURFACE.
4. **Tenant scope?** Does this slice write to tables keyed by your default tenant column (e.g. `org_id`), or to a sub-domain that uses a different tenant column (e.g. `tenant_id`)?
5. **New UI or modifies existing?** (If modifying existing: which screens? Look for likely candidates in `apps/web/src/pages/` — name them.)

Do NOT ask about things the templates force you to decide later (specific DDL columns, exact prompt text, copy deck wording). Those are Phase 3 concerns.

After answers arrive (or user defers), write `docs/slices/{{SLICE_ID}}/INTAKE.md` capturing:
- the original ONE_LINER
- question → answer pairs
- explicit `assumptions[]` list for anything inferred

## Phase 2 — Foundation (01 Persona + 02 Journey, serial)

These two docs have the highest fidelity bar and drive everything else. Write them in order, serial (NOT parallel — 02 references 01).

### 2.1 Write `01-PERSONA.md`

Do NOT invent a persona from thin air. Search existing docs for a matching persona first: