implement

name

implement

description

Build the feature or fix spec'd in blueprint/context/current-feature.md, one small reviewable step at a time. Creates the branch, implements each step, shows the diff and explains it in plain English, tests, and iterates until it works. After each approved step it offers an optional commit checkpoint on the branch; the feature-level commit, merge, and logging are /complete's job. Use when the user runs /implement, or asks to build, implement, or start the current feature or fix once its spec is ready.

implement - build the current feature or fix, one reviewed step at a time

Where this sits in the workflow:

/feature or /fix  ->  [implement]  ->  /complete  ->  next
(the spec)            (build it,       (commit +
                       reviewed)        merge + log)

/feature (or /fix) wrote the spec to blueprint/context/current-feature.md and stopped. This skill turns that spec into code, following the build loop in blueprint/context/ai-interaction.md, without vibe coding: small steps, a visible diff plus a plain-English explanation for each, testing, and iteration until it works, all behind your approval. It builds on a branch and offers an optional commit checkpoint after each step; the feature-level commit, merging, and logging are /complete's job.

Before you start

Read blueprint/context/current-feature.md. If it has no real spec (still the stub, or its status is already complete), stop and tell the user to run /feature (for a planned feature) or /fix (for an ad-hoc bug or change) first. Pull the conventions from blueprint/context/coding-standards.md and the data model from blueprint/context/project-overview.md so the code matches them.

If the spec's Design reference points at prototypes/*.html, those mockups are the visual target - build components to match them, and treat prototypes/theme.css as the token source (the spec's first step ports it into the app's global stylesheet before the components are built).

Resuming? If the spec already has some build steps checked off (- [x]), this feature was started earlier and interrupted (often a cleared context). The spec and its ticked steps are files, so pick up where it left off: read which steps are done, check the git branch and git status/log to see what is committed and what is still in the working tree, then continue from the first unchecked step instead of starting over. No separate save/load is needed - the project instructions load current-feature.md every session.

Step 1 - branch

Create and check out a branch named from the spec: feature/<name> for a feature, fix/<name> for a fix. If the project isn't a git repo yet, say so and ask the user to run git init first; the loop needs branches. On resume, the branch already exists - check it out instead of creating a new one.

Step 2 - build one step, review, iterate, checkpoint

Work through the spec's build steps in order, one at a time. For each step:

1. Implement just that step: the smallest change that satisfies its "done when."

2. Show the diff, not whole files.

3. Explain it, and prove it. Give a short summary: what the step delivered, one line per changed file on what it does and why, then confirm the step's "done when" is met with evidence (build output, a screenshot, a passing assertion). This summary is the comprehension gate, so keep it concrete, not ceremonial. Include a short How to try it note when the step has a manual path: the command, URL, click, endpoint, or output the user can check.

4. Build, and test if the project tests. Run the project's build command (see Commands in AGENTS.md). If the project declares a test command in AGENTS.md (the opt-in switch), run it: a step that adds logic must ship a passing test in the same diff, and the suite must be green before the step is approved (see the Testing gate in coding-standards.md); UI and integration-only steps ride on screenshot plus build evidence. For UI or integration done-whens, prefer Playwright when it is already installed or declared in AGENTS.md; do not add it silently for an unrelated feature. Create focused test files next to the source they cover, per coding-standards.md. Never install a runner mid-step unless the current spec is explicitly the unit-testing setup itself (for example /fix "add unit testing"). If a step surfaces non-trivial logic the spec didn't foresee, add a focused test then, or note why not. When a step's done-when is behavioral (a click, a download, a flow across screens), run /check to prove it against the running app rather than eyeballing it.

5. Iterate until it works. If it fails or the user wants changes, revise the step (re-prompt or hand-edit the code), show the updated diff, and re-test. Repeat until it works and the user approves. Nothing is committed until the user is happy with the step.

6. Mark it done, then prompt to move on. Once the step is approved, check that step off (- [x]) in blueprint/context/current-feature.md so progress survives a context clear. Then offer a short choice, noting that checkpoints are optional since /complete makes the real feature-level commit. Use the current tool's short user-input prompt when available; when you've just produced a long block to read (a deep explanation, a big walk-through), ask in plain text instead, so the prompt doesn't cover what the user is still reading:

Never batch the whole thing into one diff. If a step's diff is too big to read, split it. Build and tests must pass before any commit.

Step 3 - hand off to /complete

When every step is built and the build and tests pass (committed as checkpoints or not), stop. Tell the user it's done on its branch and to run /complete, which makes the one feature-level commit, logs it (archive, check off if it's a feature, reset), and merges. Include a short How to try it section, or tell the user to run /try for the full manual walkthrough. This skill does not touch main.

Rules

• One small step per diff; the user reviews and approves each before any commit.
• Explain every change in plain English. Understanding the code is the point.
• Iterate on the branch until each step works; never commit code the user hasn't approved.
• Follow blueprint/context/coding-standards.md (server vs client, scope user-owned queries by the authenticated user id, validate inputs, and so on).
• Build only what the spec says. If the spec is wrong or thin, stop and fix the spec first, do not improvise.
• Per-step commits are optional checkpoints. The feature-level commit, the merge, and any push are /complete's job.

Formatting

Format the output to match the project's conventions in blueprint/context/ai-interaction.md: concise, scannable markdown, with lists for enumerations and tables for matrices rather than dense paragraphs.