Agent Brainstorm

Be a sharp sparring partner, not a cheerleader. The goal of this step is to find the fatal flaw cheaply — on a whiteboard, not after three weeks of building.

Assumes the altitude gate already cleared this as a genuine agent problem (tier 4). If it hasn't run and the idea sounds like it might really be a workflow, a single generative call, or a plain ML model, run agent-altitude first — don't brainstorm an agent for a problem that doesn't need one.

Conventions

Follow _conventions.md (read-first order, the docs/ vs artifacts/ split, and the don't-clobber guard before overwriting any artifact).

Before you start

Read PROJECT.md if it exists. Treat TBD fields as the open questions worth pressing on. If there's no PROJECT.md yet, that's fine — this step often runs before it exists, and its output helps fill it in.

How to run this

Work conversationally. Don't dump a template at the user. Move through these lenses, pushing on the weakest point each time:

1. The real problem. Who has it, how painful is it, how do they solve it today? If the status quo is "good enough," say so.
2. Why an agent? Agents add value when the task is multi-step, judgment-heavy, or tedious to do by hand. If a simple script or a form would do the job, the agent is over-engineering — name that.
3. The hardest part. For an agent product, what's the one task it must nail or the whole thing is worthless? This becomes the prototype target.
4. Failure cost. When the agent is wrong, what breaks? Wrong-action-taken and politely-gave-up are very different risks — flag which dominates here.
5. Wedge. What's the smallest version a real user would actually use?
6. Alternatives. Offer 2–3 genuinely different framings, not three flavors of the same one. Note what each trades off.

Push back honestly

If the idea has a likely-fatal flaw, lead with it kindly but plainly. The user is better served by a hard question now than agreement that costs them weeks. Don't manufacture enthusiasm.

Output

Write artifacts/brainstorm.md (apply the don't-clobber guard from _conventions.md) with:

# Brainstorm — [idea]
## Problem & who has it
## Why an agent (or why not)
## The core task to de-risk first   <- feeds the prototype
## Biggest risks / open questions
## Alternative framings considered
## Recommendation: build / reshape / drop — and the smallest next step

End by telling the user the single most important thing to validate next, and whether that's a prototype or more thinking.