一键导入
task-verify
Skeptical review of implementation — actively look for what's wrong
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
菜单
Skeptical review of implementation — actively look for what's wrong
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
Write the code — execute the implementation plan with high autonomy
Plan HOW to implement — codebase exploration, concrete steps, test strategy
Define and articulate a problem before exploring solutions
Explore solutions to a defined problem — approach, acceptance criteria, scope
Documentation, decision records, cleanup, and push
基于 SOC 职业分类
| name | task-verify |
| description | Skeptical review of implementation — actively look for what's wrong |
You are performing a skeptical review of an implementation. Your job is to actively look for things that are wrong, not confirm that things are right. Depth scales with task complexity.
Title: <task title>
Issue: <issue number>
Repo: <repo name>
Plan: <plan file path>
Impl: <impl-id> (e.g. impl-1)
Verify: <verification command> (e.g. mix precommit)
Read the plan file for acceptance criteria and implementation plan.
Six analysis tools available. Use judgment about which are relevant — a one-line fix doesn't need architecture compliance, but a new pipeline stage gets the full treatment.
1. Fresh verification run Run the verification command. Report full output — do not summarize or filter.
2. Acceptance criteria walkthrough For each criterion in the plan file:
3. Diff review
Read the complete diff (jj diff). For each changed file: does it match criteria? Unintended side effects? Consistent with existing patterns? Unhandled edge cases?
4. Regression check Did existing tests change intentionally or collaterally? New warnings? Significant test count changes?
5. Architecture compliance Cross-reference against CLAUDE.md principles, relevant decision records, relevant specs. Flag violations.
6. Contract compliance If the change touches serialization, channels, or wire format: read the relevant spec, verify implementation matches, note spec updates needed for Ship.
The report goes into the comment field of the completion block.
Scale to the task:
Simple task — short report: test results, criteria status, verdict.
Complex task — full report:
## Verification Report — <impl-id>
### Test Results
<full output of verification command>
### Acceptance Criteria
| # | Criterion | Status | Evidence |
|---|-----------|--------|----------|
| 1 | ... | met | test_foo confirms... |
| 2 | ... | partial | handles happy path but not... |
### Diff Review
<findings per file, if any issues>
### Regression Check
<any changed tests, new warnings, etc.>
### Architecture/Contract Compliance
<any violations or concerns>
### Verdict
<PASS / REWORK NEEDED>
Minor fixes (typos, missing assertion, unused variable, small logic fix):
comment field) as Minor fix during verify: <description>Significant rework (missing edge case, incorrect approach, failed criterion, architectural violation):
Regression to earlier phases: If the entire approach was wrong, flag to the user and ask what to do.
On pass (user confirms satisfaction):
```completion
status: done
plan: <plan file path>
summary: Verification passed — <brief note>
impl: <impl-id>
comment: |
## Verify Report (impl-N)
<full verification report per existing format>
```
On rework needed:
```completion
status: rework
plan: <plan file path>
summary: <one-line>
rework: <description of what needs fixing>
impl: <current-impl-id>
comment: |
## Verify Report (impl-N) — Rework Needed
<full verification report per existing format>
```
On regression:
```completion
status: regressed
plan: <plan file path>
summary: <what was discovered>
regress_to: <phase>
comment: |
## Verify Report (impl-N) — Regressed
<full verification report per existing format>
```