| name | implement-plan |
| description | Implement a feature from a Cursor plan file produced by the CreatePlan tool. Use when the user asks to implement, build, execute, run, or start a plan - especially right after a plan was created by a planning skill like plan-github-issue. Stops before any git commit so the user can review. |
Implement Plan
Turn a finalized Cursor plan into working code. This skill picks up where planning skills (e.g. plan-github-issue) leave off, edits the files the plan calls for, and runs targeted tests. It stops before committing - the user reviews and commits.
Inputs
Parse from the user's message:
- Plan file path (optional) - a full path, or a filename under
~/.cursor/plans/ or the workspace .cursor/plans/.
- Issue / topic reference (optional) - e.g.
issue #4053, the back-to-scores plan, used to disambiguate when several recent plans exist.
If the user gives nothing, look for a plan file already referenced earlier in the conversation. If still ambiguous, list candidate plans and ask which to implement via AskQuestion.
Hard rules
- Implementation-only - do not create a new plan, and do not add tasks the plan does not list.
- Never create a git commit, push, or open a PR. Stop after tests run and report status.
- Stay within the "Affected code" / "Approach" scope of the plan. If you discover something material the plan missed, stop and surface it to the user instead of silently expanding scope.
- If plan mode is active, call
SwitchMode to agent before editing non-markdown files.
Workflow
Copy this checklist and track progress as you go:
- [ ] Step 1: Locate the plan file
- [ ] Step 2: Parse plan + todos
- [ ] Step 3: Mirror todos into TodoWrite
- [ ] Step 4: Implement each todo
- [ ] Step 5: Run targeted tests
- [ ] Step 6: Report status (no commit)
Step 1: Locate the plan file
Try in this order:
-
Path given by the user.
-
Plan file already referenced in this conversation (e.g. the most recent CreatePlan output - its path was returned as /Users/.../.cursor/plans/<name>_<hash>.plan.md).
-
Newest *.plan.md under ~/.cursor/plans/ and the workspace .cursor/plans/:
ls -t ~/.cursor/plans/*.plan.md .cursor/plans/*.plan.md 2>/dev/null | head -5
If multiple candidates look plausible, list them (filename + title from the first heading) and ask via AskQuestion which to implement.
Step 2: Parse plan + todos
Read the plan file in full. Extract:
- Affected code - the files you're allowed to touch.
- Approach - the step-by-step actions.
- Tests - the test files to run / extend (used in Step 5).
- Rollout - migrations, feature flags, i18n, docs to update.
- Out of scope - things you must NOT touch.
- The plan's
todos (from frontmatter) and the body's ## Approach numbered list.
Required sections. ## Out of scope and ## Acceptance criteria are mandatory. If either is missing from the plan, stop and report — do not proceed without a scope fence or an AC list. none is an acceptable value for Out of scope but the section header must be present.
Todos handoff. Read both the todos field from the plan's frontmatter AND the body's ## Approach numbered list. If both are empty, stop and report — there is nothing to implement against.
Step 3: Mirror todos into TodoWrite
Create one TodoWrite item per todo from the plan, preserving wording. Mark the first as in_progress. Update statuses as you progress through Step 4.
Step 4: Implement each todo
For each todo:
- Pattern check. Before the first edit in a directory you have not touched in this run, list 2-3 sibling files and skim one to learn the local convention (file layout, naming, helper usage, test style). Cite the file you used as a model in your final status report under "Pattern modeled on:". Skip this when editing a file you've already read in this run.
Read the target file(s) before editing.- Make the edit described in the plan (
StrReplace / Write / EditNotebook as appropriate). Match existing indentation and style.
- After substantive edits, call
ReadLints on the touched files and fix any lint errors you introduced (do not chase pre-existing ones).
- Mark the todo
completed and move to the next.
If the plan calls for behaviour the existing code does not support, stop and report rather than inventing scope.
Step 5: Run targeted tests
AC → test mapping first. Before running anything, list each acceptance criterion from the plan's ## Acceptance criteria and the test (or planned screenshot for UI-visible AC) that exercises it. Use the plan's ## Tests section as the canonical list of files. If any AC item has no backing test and no planned screenshot, stop and report — do not write a tautological test against the code you just wrote. Coverage must come from the AC, not from the diff.
Then run the listed test files (## Tests section) plus any spec files clearly covering the source files you changed. Detect the test runner from the repo:
- Rails / RSpec:
bundle exec rspec <files> (use bin/rspec if present).
- Node + Jest:
npx jest <files> (or the script defined in package.json).
- Node + Vitest:
npx vitest run <files>.
- Python + pytest:
pytest <files>.
Do not run the full suite unless the plan explicitly asks for it.
Retry budget. If a spec fails, fix the issue (within plan scope) and re-run. Cap retries at two fix attempts per failing spec file. After the second failed attempt, stop and surface the failing assertion with the source line. Never insert skip, pending, xit, xdescribe, or equivalent to make a failure go away.
If tests fail for reasons clearly outside the plan (pre-existing failures, infrastructure), stop and report.
Step 6: Report status
Print a short summary:
- Files changed (paths as markdown links).
- Tests run and their outcome.
- Any deviations from the plan and why.
- Next steps for the user, e.g. "review the diff, then commit when ready".
Do not run git add, git commit, git push, or gh pr create.
Output rules
- Cite changed files with markdown links to their workspace paths.
- Keep the final summary tight - 5-10 bullets max.
- Never invent additional refactors or "while we're here" cleanups.
Next step
After the status report, suggest the user invoke code-review against the diff and, for user-visible changes, verify-fix-evidence to capture browser screenshots. Do not auto-invoke them. For the orchestrated end-to-end flow, see the ship-issue skill.
