Builds a change end to end, from the issue to a reviewed draft PR. Implement, review, hunt invariants, QA in real projects, open the PR, act on the bot reviews, then hand over to a human.
83%
Does it follow best practices?
Run evals on this skill
Adds up to 20 points to the overall score
Low
Low-risk findings worth noting
Takes a change from a spec to a draft PR that is ready for a human. Steps 1 to 9 run without the user. Hold the decisions that belong to the user for step 9: a design-level fix (one that changes a rule the spec states), and a high-severity bot finding you disagree with. Invoking this skill authorizes the draft PR, the review request on it, the replies on bot threads, and the status change of the Linear issue.
Each step names the skill that does the work. Read that skill when you reach the step, not before.
Pin the spec. The problem to solve: an issue the user named (Linear, GitHub), a prompt, or a problem from earlier in this conversation. Write it to a spec file outside the checkout, with the decisions made so far, and give that file as the spec to the reviews in steps 3, 4 and 7: a subagent cannot read this conversation. When the spec comes from a Linear issue, set that issue to In Progress now, before any other work. Branch from origin/next and record the base SHA: origin/next moves during a session.
Implement. Follow AGENTS.md and the guidelines it points to. Run the focused tests and cd code && yarn fmt:write, then commit. The reviews compare commits, so they do not see uncommitted work; commit each later fix too, and step 6 shapes the history. Done when the focused tests pass and the working tree is clean.
Review. Run code-review since the base SHA. Fix the valid findings, commit the fixes, and run it again until only findings judged and left on purpose remain.
Hunt. Run invariants-hunt since the base SHA. Fix what survives the evidence gate. Hold the design-level findings for step 9.
QA in the real world. Run real-world-qa. Fix and commit the findings, and keep the evidence for the PR body.
Shape the history. Rebuild the commits as the pr skill "Commit history" section says.
Open the draft PR, and get two reviews in parallel.
open-pr, without its questions: choose the type and QA labels yourself with the heuristics of the pr skill, and skip the canary. Add no ci:* label; see Pull requests.Closes #<number> for a GitHub issue, Closes [SB-<number>](<Linear URL>) for a Linear issue, and the same form for any other tracker. Write Part of when the PR does not finish the issue.second-opinion on the PR.Act on the reviews. Wait for the Greptile review and collect its findings with check-pr. Handle each bot thread as the handle-pr-comments "Bot comments" section says: reply before you resolve. Leave each human thread to the user, unchanged. Also fix the confirmed findings of the second opinion, and push. Then list the unresolved review threads of every open PR in the stack, not only this one. Done when the only unresolved bot threads are the high-severity findings you disagree with.
Hand over. Tell the user the PR is ready for review. Give the PR link, the labels you chose, and, per reviewer (code review, hunt, real-world QA, Greptile, second opinion), what it found, what you fixed, and what you dropped and why. List the human comments that arrived. Present each held decision with a recommendation: the design-level fixes, and each high-severity bot finding you disagree with, with the bot's reasoning and yours side by side.
The human leads from here:
pr skill "Commit history" section says why. After a big change, run steps 3 to 5 again.ready-for-review and name the reviewer.storybookjs/storybook
7c89155
If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.