Agent-Skills.md

Agent Skills: Implement

Implement work from a ticket, spec, or inline description; auto-detects the next item from the task folders when no argument is given.

UncategorizedID: lttr/claude-marketplace/implement

Repository

lttr/claude-marketplace
lttrLicense: MIT
21

Install this agent skill to your local

pnpm dlx add-skill https://github.com/lttr/claude-marketplace/tree/HEAD/plugins/aiwork/skills/implement

Skill Files

Browse the full folder contents for implement.

Download Skill

Loading file tree…

plugins/aiwork/skills/implement/SKILL.md

Skill Metadata

Name
implement
Description
Implement a single ticket or a small inline change in one pass, in this session.

Implement

Follows the aiwork-protocol skill. Don't enter plan mode — the ticket is the plan. One work item, one pass, this session; for a whole spec with a ticket graph use /implement-spec.

Resolve input

$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.
  • Inline description: if trivial (fits one session, obvious approach), treat it as the work item. If non-trivial, stop and suggest /to-spec.
  • Spec/PRD or task folder path: if it has a tickets/ folder, or is too big for one pass, stop and suggest /implement-spec. A small ticketless spec is a valid work item: take it.

Do

  1. Set ticket status: in-progress (if working from a ticket).
  2. Implement. Use /tdd where possible, at the seams recorded in the spec's Testing Decisions section.
  3. Tidy the diff yourself — naming, duplication, dead scaffolding. Not /simplify: its reviewers would rediscover context you already have. Skip for a mechanical edit.
  4. Run the project's check command (see its scripts). Green before anything below.
  5. Run /verify <ticket-path> so it verifies the ticket's acceptance criteria, not just the diff. Check off - [ ] → - [x] for each criterion it passed.
  6. Run /code-review low --fix. Re-run the checks if it changed code.
  7. Set ticket status: done and verified: to the passes that ran (checks, behaviour, ux, review), then commit. Do not ask. Report what was done and anything left open.

When the work item is a whole spec rather than a ticket, the same upkeep applies to the spec's own status if it carries one: in-progress at the start, agent-done at the end. Never done: that is the human's sign-off once they have checked the feature, the UX and the code. Say so when you report. A single ticket finishing says nothing about its spec, so leave the spec's status alone in that case.

Stopping is hooked: the plugin's verified-gate holds the turn when a ticket claims done without an on-app pass (behaviour, ux or human in verified:) or with unticked acceptance criteria. Set the frontmatter because the passes ran, never to get past the hook. If a pass genuinely cannot run, leave the ticket in-progress and say why in implementation-notes.md. See /aiwork:implement-spec for details.

Record a note in the task folder's implementation-notes.md per aiwork-protocol when the reader must act on it: work only a human can finish, anything left unverified, a spec ambiguity you resolved, a deliberate deviation. A few sentences, never a log of what you built. Nothing to note is the normal case.