Claude Skill

evidence-first-review

Review an agent's finished work by reading the evidence before its summary, naming what each check proves and what it misses, pulling out the decisions it made silently, and trying to falsify the result before accepting it. Use when a MiniMax Code session, subagent or cloud agent

LLM Mart · 0 points · 0 views 0 listing impressions 0 install-command copies
Download minimax-ai-mcode-stack-stacks_ronny_skills_evidence-first-review-d520b2b.zip · 1 KB

Install

skills CLI npx skills add https://github.com/MiniMax-AI/mcode-stack/tree/main/stacks/ronny/skills/evidence-first-review
Claude Code claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install minimax-ai-mcode-stack@llmmart
Git git clone https://github.com/MiniMax-AI/mcode-stack.git

The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole minimax-ai/mcode-stack collection as a plugin from our marketplace. Git is the plain clone.

Skill manifest

Evidence-First Review

"All tests pass, ready to merge" is a claim, not evidence. We read what the agent actually ran before we read what it says about it. The bar for accepting the work: we could defend it to the person who will maintain it, and we can say what our checks would miss.

This skill is about owning the result. For a clean-context code review of the diff itself, pair it with independent-review.

1. Evidence before summary

Open the transcript or tool log, not the final message. For each command that supports "done": did it run on the final code, did it exit 0, does the output say what the summary says? Typical gaps we find:

  • tests run before the last edit;
  • a filter like -t compaction that skipped the suite that matters;
  • "fixed" with no reproduction before the fix;
  • a build that passed only because a step was skipped.

2. For each green check, one property it proves and one it does not

Check Proves Does not prove
New unit test passes the function returns X for this input the caller passes that input; the test was written with the same wrong assumption
pnpm typecheck types line up runtime behaviour
E2E with a mock model the loop wires up anything about real model output

If we cannot fill the right column, we do not understand the check yet.

3. Ask for the decisions list

Ask the agent: which defaults did you pick, what did you leave out of scope, and where did you reinterpret the requirement because two parts clashed? These are the choices that end up in production without anyone deciding them. Each one is accepted, changed or sent back.

4. Try to break it

Ask: what would have to be true for this to be wrong? Then check that one thing ourselves. For analysis and docs, open two of the cited sources and recompute one number the conclusion depends on.

5. Make sure we understand it

  • Pick a few lines of the diff at random and explain them without help.
  • For a large change, ask the agent for a short report that ends in three questions about the change; we answer them before merging.
  • For a contested call, have a fresh subagent argue the other side, then decide ourselves.
  • Last question to a fresh context: "What did I miss?"

6. Decide

  • Accept: the evidence matches, decisions are reviewed, we can name the gap.
  • Ask for more: missing evidence or an unexplained decision.
  • Get a human review: high stakes and we cannot say what the checks miss.

Output

When we report the work to others, we lead with the evidence (commands and results), then: what was checked, one thing that was not, and the key decisions.

References

Files (mcode-stack)
  • SKILL.md 3.1 KB
    ---
    name: evidence-first-review
    description: Review an agent's finished work by reading the evidence before its summary, naming what each check proves and what it misses, pulling out the decisions it made silently, and trying to falsify the result before accepting it. Use when a MiniMax Code session, subagent or cloud agent reports a task done, before merging its PR or acting on its analysis.
    ---
    
    # Evidence-First Review
    
    "All tests pass, ready to merge" is a claim, not evidence. We read what the
    agent actually ran before we read what it says about it. The bar for accepting
    the work: we could defend it to the person who will maintain it, and we can
    say what our checks would miss.
    
    This skill is about owning the result. For a clean-context code review of the
    diff itself, pair it with `independent-review`.
    
    ## 1. Evidence before summary
    
    Open the transcript or tool log, not the final message. For each command that
    supports "done": did it run on the final code, did it exit 0, does the output
    say what the summary says? Typical gaps we find:
    
    - tests run before the last edit;
    - a filter like `-t compaction` that skipped the suite that matters;
    - "fixed" with no reproduction before the fix;
    - a build that passed only because a step was skipped.
    
    ## 2. For each green check, one property it proves and one it does not
    
    | Check | Proves | Does not prove |
    | --- | --- | --- |
    | New unit test passes | the function returns X for this input | the caller passes that input; the test was written with the same wrong assumption |
    | `pnpm typecheck` | types line up | runtime behaviour |
    | E2E with a mock model | the loop wires up | anything about real model output |
    
    If we cannot fill the right column, we do not understand the check yet.
    
    ## 3. Ask for the decisions list
    
    Ask the agent: which defaults did you pick, what did you leave out of scope,
    and where did you reinterpret the requirement because two parts clashed? These
    are the choices that end up in production without anyone deciding them. Each one
    is accepted, changed or sent back.
    
    ## 4. Try to break it
    
    Ask: what would have to be true for this to be wrong? Then check that one thing
    ourselves. For analysis and docs, open two of the cited sources and recompute
    one number the conclusion depends on.
    
    ## 5. Make sure we understand it
    
    - Pick a few lines of the diff at random and explain them without help.
    - For a large change, ask the agent for a short report that ends in three
      questions about the change; we answer them before merging.
    - For a contested call, have a fresh subagent argue the other side, then
      decide ourselves.
    - Last question to a fresh context: "What did I miss?"
    
    ## 6. Decide
    
    - **Accept**: the evidence matches, decisions are reviewed, we can name the gap.
    - **Ask for more**: missing evidence or an unexplained decision.
    - **Get a human review**: high stakes and we cannot say what the checks miss.
    
    ## Output
    
    When we report the work to others, we lead with the evidence (commands and
    results), then: what was checked, one thing that was not, and the key
    decisions.
    
    ## References
    
    - Addy Osmani, ["Critical thinking when agents do the work"](https://x.com/addyosmani/status/2117512471930167377)
    

Comments (0)

Sign in to join the conversation.

No comments yet.

Reviews (0)

No reviews yet.