| name | spec-implement |
| description | Implement one task from a spec's tasks.md, with tests, and run the gates. Use when asked to "implement a task", "build the next task", "work on spec NNN", "do task T2", or to continue implementing an existing plan. One task at a time, never two. |
Implement one task
Work from specs/NNN-name/tasks.md. Implement the next unchecked task, or
the specific one named. One task only — not two because they are related,
not "while I was in there".
Portable equivalent of the /spec-implement prompt file. Prompt files work
only in the VS Code Chat view; this skill works there and in the Copilot CLI
and the coding agent.
Scope discipline
If you find something else broken, note it and move on — or add a task for it. A
diff that does more than its task is a diff nobody can review, and it is how an
unrelated regression gets attributed to a feature.
If the task turns out to be wrong or impossible, stop and say so. Do not
silently substitute a different task.
Method
-
Read the task, the spec's acceptance criteria, and the plan's risks.
-
Open the files you will change and the tests around them. Match the
surrounding idiom — this codebase has consistent conventions and a foreign
style is a defect in itself.
-
Check current APIs with the context7 MCP server rather than writing
from memory. SQLAlchemy 2.0, Pydantic v2, React 19, Tailwind v4.
-
Write the code and its test together. Name the test for the behaviour it
protects, not the function it calls.
-
Run the gates and show the real output:
cd apps/api && uv run ruff check . && uv run mypy src && uv run pytest
cd apps/web && pnpm exec tsc --noEmit && pnpm exec vitest run
-
If you touched analytics/, a threshold, or the generator — run the golden
tests:
cd apps/api && uv run pytest -m golden -v
If one fails, do not adjust the assertion to match your output. Work out
which behaviour changed and whether that was intended. Those tests are
designed to be hard to silence.
-
Check the box in tasks.md.
Rules that will fail review
- AI or network calls in
analytics/
agent/ querying the database instead of going through analytics/- Money as
float, or a Postgres SUM() without int() coercion — that one
passes the SQLite tests and fails only in production
- Ranking or colouring on
floor_index instead of peer_index
- Any credential, any PII, any Azure OpenAI API-key setting
- Market data without
is_synthetic
Report honestly
Say what you actually did, including what failed or what you skipped.
"It should work now" is not a result — run it and paste the output. A green
summary over a broken build costs the reviewer more than it saved you.
Next
When the task list is done, verify against the acceptance criteria — see the
spec-verify skill.
