brief · diff
git:20260702.4afacee to git:20260702.82cfccf
16 added, 1 removed. Audit A to A.
What this update changes in the checks
No check changed: every fixed check gives the same answer before and after this update.
A tripped check is a label, not a verdict: read the matched text and the lines below. This week's updates that newly trip a check.
---
name: brief
description: Use when starting any new feature or product request, before writing a plan or code — turns an idea into a short approved spec by surfacing the decisions that will hurt later. Depth scales with risk tier.
---
# Brief
Part of the **Brief** phase — see the `workflow` skill for tiers and sequencing.
Turn an idea into a clear, approved spec through collaborative dialogue. The
user speaks as PM/architect (what and why); you interrogate as the senior
engineer until the 2–3 decisions that would hurt later are settled — scope,
edge cases, failure paths — then STOP. Don't design the universe.
## Tier scaling
- **T1 (trivial):** skip this skill entirely. Just do the task.
- **T2 (standard):** a few targeted questions in chat, a 3–6 line spec summary,
user says "yes, that's what I want." No document needed.
- **T3 (risky — money/auth/data/migrations):** full interrogation, written spec,
explicit approval before planning.
## The process
**Understand the context first.** Check current project state — files, docs,
recent commits, `.agent-orch/` state if present. If a question can be answered
by exploring the codebase, explore the codebase instead of asking.
+ **Read the code before you ask.** Before asking anything the code could
+ answer, gather at least one piece of evidence — grep the symbol, read the
+ file — and cite `path:line` in the question itself: not "does this touch the
+ database?" but "`orders/service.ts:142` writes to `orders` directly — new
+ column there, or a separate table?" Grounded questions get grounded answers.
+ If you truly find nothing, say what you searched and proceed as greenfield.
+
**Assess scope before refining details.** If the request describes multiple
independent subsystems ("build a platform with chat, billing, and analytics"),
flag it immediately and help decompose into sub-projects: what are the
independent pieces, how do they relate, what order? Then brief the first
sub-project. Each sub-project gets its own brief → plan → build cycle.
**Ask questions one at a time.** Multiple questions at once are bewildering.
Prefer multiple choice when possible; open-ended is fine too. For each question,
provide your recommended answer — the user corrects a recommendation faster
than they produce an answer from scratch. Focus on purpose, constraints,
success criteria, and what happens when things fail (empty states, bad input,
the payment that doesn't go through).
**T3 depth — interrogate relentlessly.** Walk down each branch of the design
tree, resolving dependencies between decisions one by one, until you reach
shared understanding. Stress-test the plan's premises, not just its details.
**Propose 2–3 approaches** with trade-offs, in plain product language. Lead
with your recommendation and why. YAGNI ruthlessly — strike features the goal
doesn't need.
**Present the design, scaled to complexity** — a few sentences for simple
work, sections for nuanced work (architecture, components, data flow, error
handling, testing). Ask after each section whether it looks right. Go back and
clarify when something doesn't make sense.
**Design for isolation and clarity.** Break the system into units with one
clear purpose each, communicating through well-defined interfaces. For each
unit: what does it do, how do you use it, what does it depend on? If you can't
change a unit's internals without breaking consumers, the boundaries need work.
Smaller, well-bounded units are also easier to build reliably — a file that
grows large is usually doing too much.
**In existing codebases:** follow established patterns. Where existing code has
problems that affect this work, include targeted improvements in the design —
but don't propose unrelated refactoring.
## Writing the spec (T3)
+ Open with a one-line **impact/effort call** (high-impact/low-effort → do now;
+ low-impact/high-effort → question the spec itself) so the reader knows whether
+ the work earns its cost. Give every requirement **testable acceptance
+ criteria** — numbered, pass/fail, no subjective language:
+
+ - ✅ "Expired orders return HTTP 410 for all 4 user roles"
+ - ❌ "The feature works correctly" / "edge cases are handled"
+
Save to `docs/specs/YYYY-MM-DD-<topic>.md` (user preferences override) and
self-review it with fresh eyes:
1. **Placeholder scan** — any "TBD"/"TODO"/vague requirements? Fix them.
2. **Internal consistency** — do sections contradict each other?
3. **Scope check** — focused enough for a single plan, or needs decomposition?
4. **Ambiguity check** — could any requirement be read two ways? Pick one,
make it explicit.
Fix issues inline, then ask the user to review the spec before proceeding.
Only move to the `plan` skill once they approve.
## Key principles
- One question at a time; multiple choice preferred; recommendation attached.
- Explore the codebase instead of asking what it can answer.
- YAGNI ruthlessly. Explore 2–3 alternatives before settling.
- Incremental validation — approval per section, not one big reveal.
- Surface failure paths early; they are where specs go wrong.
- Stop when the load-bearing decisions are settled.
- *Adapted from superpowers (MIT, © 2025 Jesse Vincent) and mattpocock-skills (MIT, © 2026 Matt Pocock); see ATTRIBUTION.md.*
+ *Adapted from superpowers (MIT, © 2025 Jesse Vincent), mattpocock-skills (MIT, © 2026 Matt Pocock), gstack (MIT, © Garry Tan), and rampstack-skills (MIT, © 2026 RampStack Co.); see ATTRIBUTION.md.*