--- name: build-ticket description: >- Builds one ready ticket through to a reviewed result with evidence, unattended unless the user asks to be in the loop: checks it's ready, builds to its design, runs its done checks, reflects, reviews, has it verified, and reports. Use when the user asks to build, implement, or pick up a ticket. disable-model-invocation: false metadata: credit: "Adapted from Matt Pocock's implement skill, github.com/mattpocock/skills (MIT, (c) 2026 Matt Pocock)" --- # Build ticket Implement the work described by a ticket, and stop with a result the user can review: built, checked, reflected on, reviewed, independently verified, and with evidence for each done check. Work unattended by default. Stop early only where this skill says to. If the user asks to be in the loop (e.g. "with me", "let's collaborate"), pause and ask them instead wherever this skill says to pick a default, make a significant decision, or carry a question into the report. Continue with their answer. ## 1. Load the ticket If the user passes a ticket reference, fetch it from the issue tracker and state its title before starting. If the reference is ambiguous, ask. Read the ticket in full, and the spec or design document its context points to. Claim the ticket in the tracker, if the tracker supports that. Skim other open tickets, and read the ones related to this one: for context, sequencing, and collisions. Don't expand scope from them. ## 2. Ground in the code Update your git branch from the repository default to ensure you're looking at the most recent code. If the current branch is the default, create a new branch for the implementation unless asked to work directly on main. Read the code the ticket touches, to know what the system does today. Read the project's glossary and the ADRs in the area, if it has them: a ready ticket doesn't restate them. Where the ticket and the code disagree, trust the code for what is true today. ## 3. Check it's ready Run the readiness check in the `/ready-ticket` skill. If any answer is no, stop: report which answers are no and what's missing, and release the claim. Unless the only no is an open blocker, also set the ticket's state to `needs-design` (see `/ticket-states`), with what's missing in a comment on the ticket. Don't fill gaps with guesses. If the user is in the loop, confirm the approach with them against the ticket's design and done checks, and name anything still assumed. Wait for their confirmation. ## 4. Build Build to the ticket's design. Keep it the minimum that meets the ticket. For anything the ticket leaves to the builder, and forks that come up while building, pick a sensible default, keep going, and note it for the report. Always stop, rather than picking a default, when: - the design can't be built as written (the code has moved on, or the design doesn't fit): don't redesign; - the ticket turns out too big for one session: don't silently trim scope; - a decision would change user-visible behaviour or be hard to reverse, and the ticket doesn't settle it. Unattended, report why, set the ticket's state (see `/ticket-states`), release the claim, and hand off: `needs-design` when it needs redesign or splitting, `needs-human` for a decision. Say why in a comment on the ticket, with the question for `needs-human`. In the loop, ask the user and continue with their answer. Run typechecking regularly, single test files regularly, and the full test suite once at the end. ## 5. Run the done checks Run each of the ticket's done checks and keep its evidence. An inconclusive result, or evidence from the wrong surface, is not a pass: say so. If a check can't be run as written and you adapt it, the adaptation is a decision: report it. ## 6. Reflect and review Apply the `/reflect` skill, then the `/review-diff` skill. Carry their open questions and design questions into the report. Re-run any done check whose evidence the changes since step 5 could affect. ## 7. Verify Comment on the ticket (see `/ticket-states`) with an `Implementer: ` line, the decisions you made, and the done checks you adapted, so the verifier checks against the ticket you built to. Record any changes to the brief as `/ticket-states` says under **Changing a brief**. Leave out your evidence and results. Then apply the `/verify-ticket` skill. - **FAIL**: reproduce each failed check yourself, and fix the ones you can confirm. Treat the fix as steps 5 and 6 do: re-run the done checks it could affect, then reflect and review it. Comment any new decisions on the ticket, then apply `/verify-ticket` once more, with a fresh verifier. If it still fails, or you dispute the verdict, set the ticket's state to `needs-human`, with the failing checks and any rebuttal in a comment. Never override a verdict. - **INCONCLUSIVE**: no fix round. Then hand off. ## 8. Hand off Don't commit. Don't close the ticket: tickets close only when the change lands on `main`. Comment on the ticket: what was built, and the done-check results. Unless step 7 set `needs-human`, set its state to `needs-review` (see `/ticket-states`). Keep the claim. Report to the user. If the verdict is INCONCLUSIVE, say so and why before anything else. 1. **Built**: the ticket's title, and what changed. 2. **Done checks**: each check with its result and its evidence, or where the evidence is. 3. **Verification**: the verdict and the verifier's model; each check's result, surface, and evidence; and what stronger verification would need. 4. **Assumptions & decisions**: defaults you picked, builder's calls you made, and done checks you adapted. 5. **Open**: questions from reflection and review, and anything unfinished. 6. **Human review**: what the ticket asks a person to look at, if anything.