Brainstorming Agent Skill

Brainstorming

Brainstorm before you build

• Before starting any creative or feature work — a new feature, added functionality, a behavioral

change, a nontrivial design decision — stop and run this workflow before writing implementation code.

• The aim: turn the request into a validated design, recorded as a spec-graph task-spec, that the user

has explicitly approved — not a guess you implement and hope lands.

• Never implement during brainstorming. If you catch yourself opening a source file to make a change

before the design is approved, stop.

Anti-pattern: "this is too small to need this"

Every request goes through this, however small it looks. A one-line config change and a new subsystem

both benefit from a few minutes of "what does the user actually want and why" — that is where wrong

assumptions get caught cheaply. Scale the *depth* to the task; never skip the workflow entirely.

The workflow

• Orient. Use the spec-graph skill's tools first — spec_grep/spec_get/spec_graph — to find

what the project already says about the area; read code second, to confirm details.

• Scope check. If the request bundles multiple independent features or subsystems, say so and

brainstorm them one at a time (or in parallel sub-sessions, the user's call) — don't blend unrelated

decisions into one task-spec.

• Open a task-spec. As soon as you understand roughly what's being asked, spec_create a

task-spec at .thinkrail/context/TASK-<slug>.md (id, title, status: draft, parent: the nearest

relevant module) to hold the design as it develops. .thinkrail/context/ is the workspace's

gitignored scratch dir (host-seeded, zero git footprint) yet stays scannable by the spec tools — the

home for every temp doc, never committed. This file is the one artifact — update it live as decisions

land; don't also keep a separate scratch doc. This works even in a project with no existing spec graph: a task-spec only needs

frontmatter id and type to be a valid spec, no pre-existing graph required — don't skip this step

just because nothing else in the project is specced yet.

• Clarify. Ask what you need via ask_user_question, composing rounds per the

asking-user-questions concept skill — read it before the first round. Resolve a full round,

update the task-spec with what you learned, and only open a new round if the answers raised a

genuinely new question. Per that concept's degradation norms, skipped questions or a host with no

UI are not blockers: record your best-guess assumptions in the task-spec, explicitly marked

unconfirmed, and continue.

• Propose approaches. Once the ask is clear, write 2-3 approaches into the task-spec with

trade-offs and a recommendation. When approaches are easiest to compare side by side, ask via a

single-select ask_user_question with each approach as an option (label = approach name, description

= its trade-off) instead of prose alone.

• Present the design. Write it into the task-spec in sections scaled to their complexity; confirm

with the user as each section lands, not only at the end.

• Self-review. Before asking for final sign-off, reread the task-spec for: placeholders/TBDs,

sections that contradict each other, scope that's actually multiple task-specs, and ambiguous

requirements — fix what you find, don't just flag it.

• Promote. When the design settles a boundary, contract, or decision that belongs in a durable

spec, fold it into the relevant module's SPEC.md now — spec_create for a new module, spec_update

for its frontmatter (draft → active as it firms up), edit for prose. Run spec_validate after

structural changes.

• Final review, then build. Ask the user to review the (now-promoted) design once more. Once

approved, implement directly against it — there is no separate plan-writing step here. Before

handing off, self-review the implementation diff the way step 7 reviewed the spec: no silent

lint/type suppressions (a gate error is a design signal — question the flagged state or dependency

before guarding it; any genuinely-needed suppression gets explicit user sign-off first), no

nontrivial derivation duplicated across files (centralize it), no rationale left as code comments

(near-zero comments: decisions and invariants go to the owning spec per the writing-specs bar;

only lint directives and rare one-line hazard notes survive), and when the change replaced a

pattern, sweep the repo for remnants of the old one. Keep the task-spec and the durable specs

honest as the code lands, and retire the task-spec once the work itself is done, not merely

once the design was promoted.

What a good task-spec looks like

• Scoped to one piece of work — if it's accreting unrelated decisions, split it.
• States the request, the decision(s) made and why, the approaches considered and why they were or

weren't picked, and anything the user explicitly deferred or declined to answer.

• Gets promoted, not copied: once a decision belongs in a module's SPEC.md, move it there and

reference it from the task-spec rather than keeping two copies that can drift.