Claude Skill

taste-to-checks

Turn a review correction we keep making on agent output into something that runs on its own, a test, lint rule, hook, CI step or project rule, so the miss does not come back next session. Use when we keep correcting agents on the same mistake (import order, naming, a forbidden AP

LLM Mart · 0 points · 0 views 0 listing impressions 0 install-command copies
Download minimax-ai-mcode-stack-stacks_ronny_skills_taste-to-checks-d520b2b.zip · 1 KB

Install

skills CLI npx skills add https://github.com/MiniMax-AI/mcode-stack/tree/main/stacks/ronny/skills/taste-to-checks
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

Taste to Checks

If we tell an agent the same thing twice, the second time is our fault. Taste that lives only in our heads has to be applied by hand in every review, and it does not scale with the number of sessions we run. We write it down where it runs by itself.

1. Catch the repeat

Signs a correction should become a check:

  • we wrote the same review comment on two agent PRs;
  • the same thing went wrong after compaction or in a new session;
  • evidence-first-review found a miss that a mechanical check would catch.

One-off quirks stay out. Only encode what has happened more than once or caused real damage.

2. Write the rule as one testable sentence

Bad: "keep the code clean". Good: "no new any in packages/core", "every new tool has a description under 200 characters", "a bug fix PR includes a test that fails on main".

If we cannot make it testable, it is a judgment call. Keep it in review.

3. Pick the strongest place it fits

Rule shape Where it goes
Code pattern to forbid lint rule or a grep in CI, with an error message that names the fix
Behaviour that must hold test, written to fail on the bad version first
Must happen on every edit or commit hook (pre-commit, or a MiniMax Code hook)
Needs context, not mechanical project rules (AGENTS.md) or a skill, as a last resort

Prefer the top of the table. Text in AGENTS.md is read and sometimes ignored; a failing check is not.

4. Prove the check works

Run it against the past mistake: check out the commit or reapply the bad diff and confirm it fails with a clear message. Then confirm it passes on current main. A check that never failed has not been shown to work.

5. Ship it small

One rule per PR, with the review comment or incident it came from linked in the description.

Output

The rule sentence, where it went, the command that showed it failing on the old mistake and passing on main, and the PR link.

References

Files (mcode-stack)
  • SKILL.md 2.5 KB
    ---
    name: taste-to-checks
    description: Turn a review correction we keep making on agent output into something that runs on its own, a test, lint rule, hook, CI step or project rule, so the miss does not come back next session. Use when we keep correcting agents on the same mistake (import order, naming, a forbidden API, a missing test), when the same review comment shows up on a second PR, or after a review finds a miss that a check could have caught.
    ---
    
    # Taste to Checks
    
    If we tell an agent the same thing twice, the second time is our fault. Taste
    that lives only in our heads has to be applied by hand in every review, and it
    does not scale with the number of sessions we run. We write it down where it
    runs by itself.
    
    ## 1. Catch the repeat
    
    Signs a correction should become a check:
    
    - we wrote the same review comment on two agent PRs;
    - the same thing went wrong after compaction or in a new session;
    - `evidence-first-review` found a miss that a mechanical check would catch.
    
    One-off quirks stay out. Only encode what has happened more than once or caused
    real damage.
    
    ## 2. Write the rule as one testable sentence
    
    Bad: "keep the code clean". Good: "no new `any` in `packages/core`", "every
    new tool has a description under 200 characters", "a bug fix PR includes a test
    that fails on main".
    
    If we cannot make it testable, it is a judgment call. Keep it in review.
    
    ## 3. Pick the strongest place it fits
    
    | Rule shape | Where it goes |
    | --- | --- |
    | Code pattern to forbid | lint rule or a grep in CI, with an error message that names the fix |
    | Behaviour that must hold | test, written to fail on the bad version first |
    | Must happen on every edit or commit | hook (pre-commit, or a MiniMax Code hook) |
    | Needs context, not mechanical | project rules (`AGENTS.md`) or a skill, as a last resort |
    
    Prefer the top of the table. Text in `AGENTS.md` is read and sometimes
    ignored; a failing check is not.
    
    ## 4. Prove the check works
    
    Run it against the past mistake: check out the commit or reapply the bad diff
    and confirm it fails with a clear message. Then confirm it passes on current
    main. A check that never failed has not been shown to work.
    
    ## 5. Ship it small
    
    One rule per PR, with the review comment or incident it came from linked in
    the description.
    
    ## Output
    
    The rule sentence, where it went, the command that showed it failing on the old
    mistake and passing on main, and the PR link.
    
    ## 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.