Agent Skills: Implement

Implement

Follows the aiwork-protocol skill. Don't enter plan mode — the input artifact is the plan.

1. Resolve input to one work item

$ARGUMENTS is one of:

• Ticket path: use directly. Verify the ticket is status: ready and every blocked_by ticket is status: done; if not, warn and ask.
• Spec/PRD path: scan its folder for a tickets/ subfolder and pick the lowest-numbered frontier ticket (status: ready, all blocked_by done). If there are no tickets, the spec/PRD itself is the work item.
• Inline feature description: if trivial (fits one session, obvious approach), treat it as the work item. If non-trivial, stop and suggest /to-spec.
• No args: find the most recently modified task folder (per aiwork-protocol conventions), then apply the same order: frontier ticket, then spec/PRD. If everything is done but no review.md exists in the task folder, go straight to Wrap-up (§4). Nothing found → tell the user and stop.

2. Clarity gate

Before touching code, check the work item is concrete enough for one run. Warn and stop for confirmation if any hold:

• Open questions, TBDs, or unresolved decisions
• Scope spans many subsystems — suggest /to-tickets instead
• Key technical choices (data model, API shape, file targets) unspecified
• Success criteria too vague to tell when "done"
• Content is contradictory, out of order, or otherwise corrupted

Warn: "<item> is underspecified because <reason>. Implement anyway?" Wait.

The gate runs once, at the top level — against the spec/PRD or work item, not per ticket. A badly specified ticket discovered mid-chain doesn't stop for confirmation: resolve it with best judgment and record the gap and resolution in implementation-notes.md.

This is the last stop. Past this gate, run unattended to the end — no confirmations, no blocking questions.

3. Execution

This session acts as orchestrator and spawns one subagent per ticket.

• Tickets exist: for each frontier ticket in blocked_by order, spawn a subagent with the ticket path and the <ticket-loop> below as its instructions. When it returns, confirm the ticket file says status: done and a commit landed. Then spawn the next. If a ticket cannot be completed (tests won't pass, blocker discovered), stop the chain and report the state — never mark it done.
• Standalone spec/PRD or trivial inline item: run the <ticket-loop> directly in this session.

1. Set ticket status: in-progress.
2. State your understanding of what needs to be done — the goal and the intended changes — before touching code. Not a question, just a statement for the record.
3. Implement. Use /tdd where possible, at the seams recorded in the spec's Testing Decisions section. Read files before modifying, follow existing patterns, don't expand scope.
4. Run /simplify — skip only when the change was a small mechanical edit.
5. Run /verify. Then confirm each acceptance criterion against actual behavior. Check off - [ ] → - [x]; set ticket status: done.
6. Commit. Do not ask.

Throughout: keep implementation-notes.md in the task folder (an aiwork-protocol artifact) as a chronological log. Two kinds of entries:

• Workflow events — one line each, always: ticket started, ticket done (with commit hash), chain stopped and why.
• Substance — recorded as it happens, not at the end: design decisions where the spec was ambiguous, intentional deviations from the spec and why, tradeoffs considered, open questions. Trivial work gets the workflow lines only.

When no ready ticket remains, proceed to Wrap-up.

4. Wrap-up

Runs once, after the last item — never per ticket. Skip if review.md exists in the task folder and no tickets finished since; otherwise save the new review as the next number (review_2.md).

1. Run the full test suite plus other project checks (lint, build).
2. Review the branch — the whole diff across all ticket sessions. /code-review low for small/routine diffs, higher effort (medium/high/xhigh) scaled to size and risk. Fix every finding. Re-run the affected tests after fixing. A finding deliberately left unfixed goes into implementation-notes.md with the reason.
3. Save the review outcome as review.md per aiwork-protocol. Its presence marks wrap-up complete.
4. Commit remaining changes. Then report: tickets completed, commits made, review outcome, anything left open.