| name | brainstorm |
| description | Narrowly-scoped problem framing for a single coding task. Use before `/quellis-architect:plan` when the task shape is fuzzy — what's the smallest version that ships value, what are the load-bearing constraints, what evidence will count as done. Triggers: brainstorm, frame, scope, what's the smallest version, how should I cut this.
|
| allowed-tools | ["Read","Grep","Glob","Write","Edit"] |
Brainstorm (coding-agent task framing)
A tight discovery pass for a single coding task. Answers what is the smallest version that ships value, and what evidence will count as done. Hands a short brainstorm note to /quellis-architect:plan, which writes the evidence contract.
This skill is deliberately narrower than deep-thought:brainstorm. That skill is for product/feature-level decisions. This one is for "I'm about to ask an agent to do a chunk of engineering — what's the right cut."
Boundaries
This skill MAY: read code, grep for patterns, ask one or two reframing questions, write a short brainstorm note.
This skill MAY NOT: write or edit code, run tests, install packages, deploy, propose the plan, write the evidence contract.
Do not write code in this skill. The whole point is to commit to nothing yet.
Common Rationalizations
| Shortcut | Why it fails | The cost |
|---|
| "The task is clear, skip framing" | Tasks dressed as clear usually carry one hidden assumption. Naming it up front beats discovering it on hour three. | Rework when the assumption breaks. |
| "Bigger cut is better — fewer round trips" | The agent loses context the larger the slice. Narrow slices fail visibly. Wide slices fail invisibly. | Silent half-completion. |
| "Brainstorming is for product, not for plumbing" | Plumbing decisions encode plumbing constraints. If you don't name them now, the agent will guess. | Wrong-shape solution. |
Phase 0: Decide whether brainstorming is needed
Skip this skill when any of these hold:
- A
docs/brainstorms/ doc already covers the task at the right cut.
- The task is a one-line bug fix with an obvious root cause.
- The user already has a clear acceptance test in mind.
In those cases, jump to /quellis-architect:plan directly.
Brainstorm when any of these hold:
- The user describes a goal but not a slice.
- Multiple plausible cuts exist, and the choice has consequences beyond this PR.
- The task touches a footgun directory (auth, migrations, scoring, payments).
Phase 1: Reframe the ask in two questions
Open with two questions, in order. Do not list more. The point is to force a small move, not a survey.
-
What's the smallest version of this that ships value?
The smallest version is rarely the user's first description. Surface it.
-
What would have to be true for you to call this done?
"Done" is the evidence contract in disguise. The answer feeds /quellis-architect:plan directly.
If the user can't answer #2 specifically, that's the brainstorm finding — the task isn't ready for an agent. Recommend they write a short acceptance test by hand first.
Phase 2: Surface the load-bearing constraints (≤ 3)
Read the immediate code surface. Grep for the patterns the task touches. Identify at most three constraints that change the shape of the solution. Examples:
- "This touches the migrations/ directory — the §2.D backfill rule applies."
- "The auth/ folder has no public API; new code must route through the existing session module."
- "There's no test infra for this path; verification will need a manual SQL query."
If more than three constraints stand out, the task is too big — recommend cutting before planning.
Phase 3: Write the brainstorm note
Write to docs/brainstorms/<YYYY-MM-DD>-<short-slug>.md (or the project's brainstorm directory override). Keep it under 30 lines. Structure:
---
title: "<task title>"
type: brainstorm
date: <YYYY-MM-DD>
status: ready-for-plan
---
# <task title>
## Smallest version
<one paragraph: what ships if we only ship this>
## What "done" looks like
<3-5 bullets of testable conditions — these become the evidence contract claims>
## Load-bearing constraints
<≤ 3 bullets, each with a code pointer>
## Open questions
<questions the user must answer before /plan; one per line; mark "BLOCKING" or "PARKED">
The brainstorm note is the input to /quellis-architect:plan. It does not need to be exhaustive; it needs to be clear about the cut.
Phase 4: Hand off
Tell the user: "Brainstorm ready at <path>. Run /quellis-architect:plan <path> next, which will write the evidence contract." Do nothing further.
What makes this Quellis-shaped
- Two questions, not twenty. The deep-thought version explores; this one commits.
- Three constraints max. More means cut harder, not plan harder.
- Always produces a written note. Verbal brainstorms vanish; the contract that follows needs an anchor.
- Stops at the contract handoff. This skill does not propose the plan. The split prevents scope drift.