Implement feature

implement-feature

New behaviour, fully integrated — not a prototype. The two expensive mistakes here are building the wrong thing and building a second copy of something that already exists in a different shape. So you earn the right to write code by first understanding the intent and feeling the ground you stand on. For a structural or performance change that keeps behaviour the same, use make-improvement; for a defect, fix-bug; for the tests, test-code; to open the PR, create-pr.

When this applies

Any task that adds behaviour a user or caller can see — a screen, a field, a route, an endpoint, a setting, a capability. Skip the ceremony only for a one-line change in a place you already know.

Write a note first

Invoke write-notes and answer this form before you implement:

• Intent: Describe exactly what a user or caller will be able to do that they can't today — and what stays the same.
• Status quo: Describe how the area behaves now and how it's built — which files/components own it, and what you saw when you ran or read it.
• Reuse: Describe the existing concept you'll build on (name the component / hook / util / endpoint and its path) and how you'll extend or compose it. If you're adding new, explain what you searched for and why nothing fit.
• Approach: Describe the thinnest end-to-end slice (data → logic → UI) and how it wires into the code already there.
• Ripple: For each area a change of this shape can touch (tests, migration, docs, i18n, a11y), describe what this one requires.
• Risks & unknowns: Where are you least confident? Describe the assumption that, if wrong, would break this — and how you'd catch it before it bites.

Gate A — before you write code

The senior move is to slow down here. Tick every box, or N/A with a reason — an unticked box means not done.

• Note — the form above is answered and written first (write-notes).
• Intent — restated in your own words; every design-changing ambiguity asked, not guessed — a wrong guess wastes the whole feature, and you keep asking the moment you hit a roadblock. Facts you lack are researched (deep-research); preferences are asked.
• Status quo — you ran or navigated the real app (or exercised the real code path) for the area you're extending; you've experienced the current behaviour, not imagined it.
• Reuse — the existing concept is found (search-codebase) and named — or you can say why none fits. Reuse → extend → generalize: a second catalog/list/form that already exists in another shape is a defect, not a feature; create new only when nothing fits, in its canonical home, under a name the next dev will search for.
• Conventions — discovered from the project's tooling and its neighbouring code (search-codebase orient), not from memory, so your code passes the gate the first time.
• Blast radius — every other site of the concept and every cross-cutting artifact enumerated (search-codebase sweep) — and the smallest, most reversible slice picked.

Build the vertical slice

• Thinnest end-to-end slice first. Make the intent work all the way through — data → logic → UI — before widening. Avoid gold-plating and speculative generality; build what was asked, well.
• Match the neighbours exactly. Mirror the structure, naming, and error handling of the files you're in. Code that looks like it was always there is the goal.
• Validate at every boundary. Treat all external input (requests, params, payloads, files) as adversarial; parse and reject at the edge.

Gate B — before you call it done

Tick every box, or N/A with a reason — an unticked box means not done.

• Slice integrated — the thin slice works end-to-end through real wiring; no dead flags or orphaned UI.
• Definition of done — walk create-pr's shared checklist now, not at PR time: green gate, tests, migration, locales, docs, accessibility.
• Sweep — every site enumerated at Gate A is changed or explicitly ruled out (search-codebase).
• Observed — the real outcome watched (test-code), never a green typecheck alone.

Then take it to a clean PR with create-pr.

Patterns

• Thin slice over big bang. Five small integrated commits that each keep the suite green beat one thousand-line drop nobody can review or revert.
• The reuse miss you can't see is the costly one. About to name a new component XList/XGrid/XPicker? Stop — search-codebase first; the project almost certainly already has the concept.