Spike

A spike is a deliberately disposable build whose only job is to answer a question. Its highest-value output is not the code — it is the production note: a written statement of what the real build needs, extracted while the problem is fully loaded in your head. A good spike with a good production note is the spec, and can be handed to any implementer (human or agent) cold.

Pick the branch

Logic spike — for "does this state machine / data model / algorithm hold up?" Build a single runnable script or minimal CLI that exercises the model with hardcoded data. No UI, no persistence unless persistence IS the question.

UI spike — for "which of these designs works?" Build 2–4 radically different variations (not one design with three paddings), all reachable from a single route with a visible toggle. Client-side state only; fake data; zero backend changes.

Rules

• Isolated: everything lives under a /prototype route or a spike/ directory. No edits to production modules. No new production dependencies.
• Timeboxed: hours, not days. A spike that grows tests and error handling has silently become an implementation — stop and restart as real work.
• Real constraints, fake data: enforce the rules that shape the design (minimum lead times, size limits, valid transitions) so the feel is honest; fake everything else.
• Never merged to production paths: archive the branch or delete it after the production note is extracted.

The production note — mandatory

At the top of the spike's main file, before finishing:

// ── Production note ─────────────────────────────────────────────
// This spike is client-side only; nothing is persisted. A real build needs:
//   - <schema deltas: columns, statuses, enums>
//   - <processes: jobs, crons, queues — and which existing path they reuse>
//   - <the decided UX rules: minimums, defaults, edge behaviors>
//   - <what was tried and REJECTED, and why>
// ────────────────────────────────────────────────────────────────

The rejected-options line matters most: it prevents the implementer from faithfully rebuilding a dead end.

Exit

Present the variations (or the logic run's output), get a decision, write the production note, and hand off — to /spec if the feature is going to be built, or to the archive if the answer was "no".