メニュー
Skeptically validate that implementation is complete and works. Heavier than the verifier agent — exercises the feature, checks integration, and confirms regression protection.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Write comprehensive tests using project conventions: AAA pattern, edge-case checklist, httpx.MockTransport, conftest fixtures, parametrize where it reads better than copy-paste.
Audit Prefab UI card builders (`build_*_ui` in `katana_mcp_server/src/katana_mcp/tools/prefab_ui.py`) for user-centric content: real names over IDs, no redundant text dumps, named helpers over ad-hoc rendering. Reports findings; does not modify code. Use when reviewing card UX or after touching `prefab_ui.py`.
Groom the Rolling Backlog (project #5) by reading current state and proposing diffs — not re-deriving the queue from scratch. Surfaces drift between issue state and project status, missing classification, and stale priorities. Asks per-category confirmation before mutating the board.
Quick pre-flight validation of staged and unstaged changes before committing. Lighter than the verifier/code-reviewer agents — focuses on catching common mistakes in the current changeset.
Scan editable code for tech debt, anti-patterns, and improvement opportunities across 5 categories. Reports findings; the code-modernizer agent applies fixes.
Open a PR for the current feature branch — self-review the diff, organize commits, push, create the PR, wait for CI and review, then address feedback. Use when implementation is complete and ready for review.
| name | verify |
| description | Skeptically validate that implementation is complete and works. Heavier than the verifier agent — exercises the feature, checks integration, and confirms regression protection. |
| allowed-tools | ["Read","Grep","Glob","Bash(uv run poe agent-check)","Bash(uv run poe test)","Bash(uv run poe test-coverage)","Bash(git diff *)","Bash(git status)"] |
Skeptical end-to-end check: not just "compiles green" but "actually works."
Before any infrastructure check, use what was implemented:
This catches "compiles but doesn't work" failures the test suite missed.
uv run poe agent-check # format + lint + type check
uv run poe test # tests
ALL green, no new warnings.
For each new public function/class, run LSP findReferences. Zero references means something is unfinished — flag it.
api/**/*.py, models/**/*.py, client.py not manually editeduv run poe test-coverage
Core logic ≥87%. New code has both success and error path coverage.
## Verification Report
### Status: [PASS | FAIL]
### Verified
- [item]: [evidence]
### Failed
- [item]: [what's wrong + fix]
### Recommendations
- [improvements noticed]
CLAUDE.md or another doc misled the implementation, fix the doc as part of verification. The goal is that future work doesn't repeat this failure.verifier agent — fast mechanical gate (this skill is the heavier audit)pr-preparer agent — process readiness (commits, ADRs, help resource sync)code-reviewer agent — design review (this skill is correctness, not design)