Brainstorm Before Building
Ask one question at a time, and treat every approval as the human's to give — never inferred from silence (collaboration, when you reach a gate).
If the codebase map exists, read .planning/power/context/project.md and
.planning/power/context/domain.md first (context defines them).
They tell you the architecture you are extending and the words this code already uses —
proposing Patient where the code says Beneficiary produces a design nobody can map onto the
repository. The map informs; it never decides. Where it disagrees with the code, the code is
right and the map is stale. Read artifacts before writing
spec.md or state.md.
Step 1 — Classify, out loud, before the first question
Say which path this is and why, in one sentence, before you ask anything. The classification determines how much process follows, so hiding it hides the decision.
| Path | What it is | Where it ends |
|---|---|---|
| Spike | A feasibility question. You do not know if something is possible. | A recommendation. Any code is labelled throwaway. |
| Bounded | A well-scoped change to a flow that already exists in this repository. | Agreement in chat. No spec file, no plan. |
| Architectural | New systems, new subsystems, interface changes, anything you cannot fully picture yet. | An approved spec.md, then power-plan. |
Bounded means you can read the flow you are about to change. Understanding what kind of application it is does not make a change bounded.
The ratchet. When torn between two paths, take the heavier one. Hidden complexity found mid-task moves you up a path; nothing ever moves you down. A bounded change that turns out to need a new interface becomes architectural at that moment, and you say so.
Spike
State the question and a two-or-three-sentence probe plan. Get a nod. Investigate the cheapest way that actually answers it. Report what you found and what you recommend. Say explicitly that any code written is throwaway.
Bounded
Ask what you need, one question at a time. Read the existing flow. Present a short design in the conversation — no file. Then stop. Implement only after an explicit yes.
Architectural
Explore the context: read the code the change touches, the tests around it, the conventions it must match.
Ask questions one at a time. A numbered list of six gets answered like a form.
Present two or three approaches with real trade-offs and a recommendation. "It depends" is not a recommendation.
Walk the design section by section, taking approval per section, so a disagreement costs one section rather than the whole thing.
Write
.planning/power/features/<slug>/spec.md:# <Feature> — Design Status: DRAFT Source: <requirement or conversation> Updated: <ISO date> ## Problem ## Approach ## Decisions ## Interfaces ## Constraints ## Out of scope ## Acceptance criteria ## RisksEvery decision gets Decision / Options / Choice / Why / Trade-off / Reversible. Constraints carry exact values — timeouts, limits, formats, names — because the plan will copy them verbatim and an implementer will only ever see the copy.
Self-review the spec before showing it: does it cover the whole problem; does it contain any placeholder or "TBD"; is it internally consistent; is anything ambiguous enough that two engineers would build different things?
Gate. Present the spec and wait. On approval, set exactly one
Status: APPROVEDfield and record the gate instate.md.
The only exit
On the architectural path, the single next skill is power-plan. Not an implementation skill,
not a UI skill, not "let me just start the first file". If you find yourself reaching for a
framework, you have skipped the gate.