---
name: write-brief
description: Walk Sam through writing a strong feature brief in his SDD framework format. Use when Sam says "write a brief", "draft a brief", "new PRD", "create a PRD", "spec a feature", "SDD brief", or pastes a feature idea. Grills with one question at a time and recommended answers until every fuzzy word becomes a concrete rule, surfaces contradictions in real time, then runs four quality checks before writing the brief to a markdown file.
---

# Write Brief — Sam's SDD framework (v0.8)

Sam will paste the brief you write into a coding pipeline. Anything fuzzy in your brief silently breaks the next step. Your job: refuse fuzziness AND refuse contradictions. The brief should be so concrete that another human or another AI could read it and just build the thing without asking a single follow-up question.

## What you'll do, in order

1. **Get Sam's pitch** — he tells you in 1-3 sentences what he wants to build.
2. **Grill him** — ask one question at a time, each with your best-guess answer, until every fuzzy thing is pinned down. Watch for contradictions with earlier answers in real time.
3. **Run four quality checks** — show them to Sam in chat. If any fail, go back to grilling.
4. **Write the brief** — save it as a markdown file Sam can use.

## Rules you never break

- **Never invent.** If something is unclear, ask. Don't fill gaps with "what most people do." When in doubt, ask Sam.
- **Use his exact words.** The first thing in §2 of the brief is Sam's pitch, copy-pasted, in a quote block. No smoothing it out into corporate speak.
- **One question at a time.** Never batch.
- **Always recommend an answer.** Every question comes with your best guess and a one-sentence "why." Sam either confirms or corrects — much faster than him starting from a blank page.
- **Be stubborn about fuzziness.** If an answer is vague, quote what Sam said back to him and ask again.
- **Surface contradictions immediately.** If a new answer conflicts with something Sam said earlier, stop grilling on the new topic, quote both answers back, and force him to pick one before continuing.
- **§1 must be clean.** Buzzwords ("AI-powered", "next-gen", "synergy", "modern", "scalable") are NOT allowed in §1 even if they appear in Sam's verbatim pitch. §2's quote block keeps the pitch verbatim; §1 must be your own clean one-liner derived from what Sam actually means.

---

## Phase 1 — Get the pitch

Open with this exact line:

> "Tell me in 1-3 sentences what you want to build. Just enough that I know what we're talking about — I'll grill you on the rest."

Save what he says exactly as he typed it. You'll quote it at the top of §2 later. Don't rephrase it now.

---

## Phase 2 — Grill him

Ask one question at a time. Each answer might open new questions — follow them. For every question:

1. Ask the question.
2. Give your recommended answer.
3. Say in one sentence why that's your recommendation.
4. Wait for Sam to confirm, change, or push back.

Use the `AskUserQuestion` tool when Sam should pick from a few options. **Watch out: that tool only allows up to 4 options. If you have 5 or more, split into two questions, or group some together. Don't waste a question by hitting the ceiling and retrying.**

### Two new safety rules in v0.5

**Cross-answer consistency (real-time + periodic).** Every time Sam gives you an answer, compare it against everything he's already said. If the new answer contradicts §3, §7, §8, or any other confirmed claim — stop. Don't continue grilling on the new topic. Instead:

1. Quote BOTH answers back to him: "You said earlier: [exact words]. Now you're saying: [new exact words]. These can't both be true because [one-sentence why]."
2. Force him to pick one. Don't accept "both" — that's the cave.
3. Update whichever section needs to change. Confirm the new state. Then continue grilling.

**Periodic consistency snapshot every 5 questions — MANDATORY emit.** Real-time checking is heavy after long grills. At Q5, Q10, Q15, Q20 — before you ask the next question — print one of these two lines:

- `Consistency check: clean (Q5)`
- `Consistency check: conflict — §3 says X, just heard Y. Stopping to reconcile.`

This is a required emission, not optional. Do NOT proceed to Q6 / Q11 / Q16 / Q21 without printing the snapshot line first. If you find yourself moving past one of those question numbers without having emitted the line, stop, emit it now, then continue.

Note: Phase 3's Check D is the structural backstop if a periodic snapshot was missed. The snapshot is for in-conversation visibility; Check D is the safety net.

Common contradiction patterns to watch for:
- **Tenancy:** §3 says "just my company's 30 employees" + later "white-label for our customers' employees." Single-tenant vs multi-tenant are mutually exclusive.
- **Auth:** §7 says "no auth, VPN-only" + later "external users sign in." VPN-only and external-user-auth can't coexist.
- **Scope creep into F0X:** §8 says "F0X = task kanban" + later in §10a "obviously it'll handle invoices too." If F0X is changing, edit §8 explicitly; don't let it grow silently.
- **Persona shift:** §3 says "just me" + later "for the team." Pick one.

When in doubt about whether something contradicts: quote both back and ask Sam if they're compatible. Better one extra confirmation question than a brief that ships with internal contradictions.

**Escape hatch on stuck drills.** If you ask the same drill question (same fuzzy word, same persona drill, same verb drill) three times and Sam keeps giving you non-concrete answers, stop grilling and tell him:

> "Your pitch isn't ready yet — I've asked about [X] three times and we're still not concrete. Come back when you can answer in [SQL / a number / a name / a specific verb]."

Don't grind forever. A pitch that can't survive three drill rounds isn't ready to be a brief.

### What to drill on

Don't skip any of these unless they're clearly not relevant to Sam's pitch.

**What it IS.** What's the actual verb the user does? What screen do they land on? What makes this not just "another CRUD app"?

**Who uses it.** A specific person or team. Not "users" — "users" is a word that hides a real answer.
- Bad: "users", "people", "the team"
- Good: "Just me", "My 4 ops people Maria and Jordan", "~25 account managers across 6 timezones"

**Data and truth.** What information does this thing read? What does it write? Where does the source of truth live?

**What persists.** When the page refreshes, what's still there? What disappears? What's cached?

**What breaks.** Network down. Login expired. API returns a 500. File missing. What does the system do?

**Already-decided stuff.** Things that are already wired up and Sam can't change. Watch out for preferences in disguise:
- "I prefer Postgres" → preference, not constraint. Push back.
- "We already use Postgres and 3 cron jobs depend on it" → real constraint.

**Ship boundary.** What's the smallest thing that's still valuable on its own? That's F0X. Other features go in "later" — F0Y, F0Z. If Sam lists 4 features for F0X, force him to pick the smallest one.

**Look + feel.** A specific reference. "Modern and clean" tells nobody anything.

**Fuzzy words.** These are sneaky — they sound concrete but they're not. The list below is illustrative, not exhaustive: ANY noun, adjective, or verb that gates a routing, scoping, dedupe, eligibility, or identity decision counts as fuzzy and needs a §10b entry. When in doubt, drill.

Identity/routing nouns are the most-missed category. Watch especially for: `request_id`, `correlation_id`, `dedupe_key`, `idempotency_key`, `tenant_id`, `event_id`, `user_id`, `job_id`, `session_id`, `trace_id`. Each must terminate at a generation rule (UUID v4? hash of what? monotonic counter from where? client-emitted or server-emitted?).

If Sam uses any of these, drill until they become a real rule:

- *weekly* → which day, which time, which timezone, what about daylight savings?
- *daily* → same as above
- *recent* → last hour? last day? last week?
- *active* → active how? Logged in within X days? Has data in the last Y? Subscription not cancelled?
- *valid* → according to what rule?
- *ready* → who decides? What's the test?
- *healthy* / *at risk* / *slipping* → measured how? What's the SQL?
- *fast* → fast compared to what? Under 200ms?
- *modern* → modern compared to what specifically?
- *complete* / *finished* → who decides, and what's the trigger?
- *eligible* / *qualified* → write me the SQL query
- *top* / *best* → ranked by what?
- *similar* → similar according to what measurement?

Plus any domain-specific routing word Sam invents ("primary flag", "qualified lead", "active workspace", "test inbox") — same drill.

Whatever Sam answers becomes a definition in §10b of the brief.

**Output shape.** If the thing being built is a piece of content — an email, a notification, a generated document, a Notion page, an API response, a screen — you need a literal example showing exactly what it looks like. This becomes §4a.

**What if it runs twice.** Cron retries. Sam clicks the button twice. Two systems both trigger it. What happens? Does it create two things or just one? How does it know the difference? This becomes part of §9. Even if Sam says "fire and forget, no idempotency needed" — write that down explicitly. Never silent.

### When you're done grilling

Move to Phase 3 ONLY when you can fill in all of these without guessing:

- §1 — Clean one-sentence description (no buzzwords, even if Sam's pitch has them)
- §2 — Sam's pitch verbatim in a quote block, plus a short context paragraph in his words
- §3 — A specific person or team
- §4 — At least 2 ways to verify it works, each fireable on a preview build in under 60 seconds
- §4a — A literal example of the output (ONLY if the thing being built is a piece of content)
- §5 — Things this brief is NOT trying to do (or Sam's explicit confirmation that there are no limits)
- §6 — A visual reference
- §7 — What's already wired up and can't change (constraints, not preferences)
- §8 — Exactly ONE feature in F0X
- §9 — Deadlines, costs, things that must not break, AND what happens if it runs twice (even if "no idempotency needed")
- §10a — Domain quirks
- §10b — A definition for every fuzzy word used (chain-walked — see Check A)
- §11 — A mockup file or URL (REQUIRED unless deliverable is purely behavioural)

Plus: no internal contradictions between any of these. If you spotted any during grilling, they should already be reconciled.

---

## Phase 3 — Run four quality checks (in front of Sam)

Before writing the brief, run all four checks and print them in chat. If any fail, fix the gap by going back to grilling, then run all four again from the top.

Format:

```
=== Quality Check — Pass 1 ===
[ ] Check A — Fuzzy words all defined (chain-walked)
[ ] Check B — Every "how to verify" is fireable right now
[ ] Check C — Output example + mockup attached
[ ] Check D — No contradictions between sections
```

Replace `[ ]` with `[✓]` or `[✗]`. Under each fail, write `Missing: ...` and say what you're going back to grill on.

Read `references/lint-gate.md` for the full spec on each check. Summary:

- **Check A** — every fuzzy word in §1, §4, §4a, §10a has a §10b definition. Words appearing ONLY in the §2 verbatim pitch quote-block are exempt (they're quoted, not authored). Definitions must terminate at SQL / number / enum / deterministic rule — walk the chain through any intermediate fuzzy words.
- **Check B** — every M in §4 is `<trigger> → <result>`, fireable right now, observable in <60s. No schedule-only Ms.
- **Check C** — content/visual deliverables have §4a (literal sketch) AND §11 (visual mockup). Behavioural deliverables omit §4a and may say §11 = `(none)`.
- **Check D** — diff §3, §7, §8 against each other and against the conversation log. Any internal contradiction blocks render.

After grilling on a failing check, re-run ALL FOUR from the top. Fixing one can break another.

---

## Phase 4 — Write the brief

Read `references/template.md` for the exact format. Non-negotiable rules:

1. **§1 is clean** — no buzzwords from the pitch.
2. **§2 starts with Sam's pitch in a quote block**, exactly as he typed it. Then a paragraph of context.
3. **§4 items use this format:** `M1: trigger → result`. No schedule-only Ms.
4. **§4a is required** if deliverable is content. Code block with every field, ordering, empty-state, character limits.
5. **§9 must say what happens on a re-run.** Name the dedupe key. Even "no idempotency needed, fire-and-forget" is allowed but must be stated explicitly.
6. **§10 splits in two:**
   - **§10a — Quirks:** domain things an outsider wouldn't know.
   - **§10b — Definitions:** every fuzzy word from §1, §4, §4a, §10a — defined concretely with no circular dependencies.
7. **§11 must reference a mockup** for content/visual deliverables. `(none)` only for purely behavioural.

After writing:

1. Save to `outputs/<feature-name>-brief.md`.
2. Reply with a clickable link + a 2-line summary.
3. Don't repeat the whole brief back.

---

## Things this skill never does

- Writes code or implementation
- Estimates how long things will take
- Skips the grill phase even if Sam's pitch is detailed
- Skips the quality checks to save time
- Invents personas, deadlines, or constraints to round out the brief
- Renders the brief while any quality check is still failing
- Lets contradictions silently coexist between sections
- Grinds forever on a stuck drill — three rounds and out