3be9b17, the version that was reviewed and scanned here. Newer commits reach this catalog only after a new review. License: MIT, by Off Grid AI.
Learn from every review
The code-check skill is only as good as the mistakes it has learned to catch. Every review is a free lesson. This skill decides which lessons are worth keeping, and folds those, and only those, back into code-check so the next PR starts from a higher bar. The discipline is as much about what you REFUSE to add as what you add: a bloated standard is an ignored standard.
Where the standard lives. code-check/SKILL.md (its PLAN, AUDIT and Standards sections) and its provenance log code-check/LEDGER.md, wherever your agent installed the skill (for example .claude/skills/, .agents/skills/, or the project's own skills folder). If the team keeps its standard elsewhere (a CONTRIBUTING.md, an engineering handbook), fold lessons there instead and keep the same ledger beside it.
Input. The review comments for a PR: from safe-to-merge Phase 2, or gh pr view <n> --comments plus gh api repos/{owner}/{repo}/pulls/{n}/comments, or a set the user pastes. Include your own AUDIT findings and any bug that slipped past the test suite: a shipped regression is the highest-signal lesson there is ("why did the suite stay green?").
Step 1: For each comment, classify GENERIC or SPECIFIC
Promote a lesson into the standard only if all five hold:
- Recurs. It is a class of mistake, not this feature's domain bug. The test: "Would this exact comment be valid on an unrelated PR, in another repo?" If it only makes sense for this feature, it is SPECIFIC: discard.
- Maps to a section. It sharpens an existing rule, tell or anti-pattern in one Standards section (Separation of concerns, Avoid local state, SOLID, SSOT / DRY, One seam across platforms, Testing, Design, Copy, Branch / PR / honesty, Mechanical gates), or it shows a genuinely missing section. Name the section.
- Detectable. You can state it as a rule with a tell: a concrete signal a reviewer or a grep can catch next time (
kind ===, a hardcoded hex color,expect(x).toHaveBeenCalled(), a duplicated constant). "Write better code" is not a tell: discard. - Not already covered. Search the current skill. If the rule is already there, do NOT duplicate it; at most tighten the existing wording. Most near-misses land here.
- Portable. It belongs in the cross-repo standard. Repo-specific facts (a path, a command, a token name) go to that repo's agent instructions (
CLAUDE.md,AGENTS.md,.cursor/rules, or the equivalent), not here. Anything a formatter or linter already enforces is not a lesson: it is a gate config (Mechanical gates).
Discard (do not touch the skill) when: it is a one-off domain bug; a style preference the formatter or linter owns; project-specific (it goes to the repo's agent instructions); already covered; or a nit with no general signal. Discarding is the common, correct outcome: most comments are SPECIFIC.
Output a short table: each comment, GENERIC or SPECIFIC, and for a generic one the target section and the tell.
Step 2: Fold the GENERIC lessons in, minimally
For each GENERIC lesson:
- Prefer sharpening an existing bullet over adding a new one. Add the missing tell or anti-pattern to the rule that already exists. Add a new bullet only when no existing rule covers it. Add a new section only when a whole dimension is missing (rare; justify it).
- Write in the skill's voice: an imperative rule, the tell, and optionally the anti-pattern. Terse. Match the density of the text around it. No preamble, no hedging.
- Keep it tight. The bar for this skill is high signal, no bloat. If your addition makes a section longer without making it sharper, cut it back. One vivid tell beats three paragraphs.
- One section per lesson. If a comment spans two sections, split it.
Step 3: Show the diff, apply, log
- Show the exact before and after edit to
code-check/SKILL.mdfor each accepted lesson. Keep edits surgical (a tightened sentence, an added tell); never rewrite a section. - Apply the edits.
- Log provenance in
code-check/LEDGER.md, one line per accepted lesson:YYYY-MM-DD · <repo> PR #<n> · <section> · <the one-line lesson>. The ledger is the audit trail: it lets anyone see why a rule exists and revert a bad addition. Use absolute dates. Keep only the one-liner in the ledger; the content lives in the skill. - Report: N comments reviewed, M generic (folded into which sections), N minus M specific (discarded, with why in one phrase). If M is 0, say so plainly. A review with no general lesson is normal and honest; do not invent one to justify an edit.
Guardrails
- The skill must not only grow. When you add, look for a rule the new tell makes redundant, and merge them. Net size stays roughly flat; density rises.
- Never weaken a rule to match a shipped mistake. Lessons raise the bar, never lower it (as with coverage: raise, never lower).
- A shipped regression changes the test class, not just the code. If the lesson is "the suite stayed green while the feature broke", the edit is a sharper Testing tell about the missing corner or seam. Then search the codebase for tests of the same shape.
- Don't over-fit to one loud reviewer. One bot's style preference is not a standard. The "would it recur elsewhere?" test must pass on its own merits.
Learn from every review
The code-check skill is only as good as the mistakes it has learned to catch. Every review is a free lesson. This skill decides which lessons are worth keeping, and folds those, and only those, back into code-check so the next PR starts from a higher bar. The discipline is as much about what you REFUSE to add as what you add: a bloated standard is an ignored standard.
Where the standard lives. code-check/SKILL.md (its PLAN, AUDIT and Standards sections) and its provenance log code-check/LEDGER.md, wherever your agent installed the skill (for example .claude/skills/, .agents/skills/, or the project's own skills folder). If the team keeps its standard elsewhere (a CONTRIBUTING.md, an engineering handbook), fold lessons there instead and keep the same ledger beside it.
Input. The review comments for a PR: from safe-to-merge Phase 2, or gh pr view <n> --comments plus gh api repos/{owner}/{repo}/pulls/{n}/comments, or a set the user pastes. Include your own AUDIT findings and any bug that slipped past the test suite: a shipped regression is the highest-signal lesson there is ("why did the suite stay green?").
Step 1: For each comment, classify GENERIC or SPECIFIC
Promote a lesson into the standard only if all five hold:
- Recurs. It is a class of mistake, not this feature's domain bug. The test: "Would this exact comment be valid on an unrelated PR, in another repo?" If it only makes sense for this feature, it is SPECIFIC: discard.
- Maps to a section. It sharpens an existing rule, tell or anti-pattern in one Standards section (Separation of concerns, Avoid local state, SOLID, SSOT / DRY, One seam across platforms, Testing, Design, Copy, Branch / PR / honesty, Mechanical gates), or it shows a genuinely missing section. Name the section.
- Detectable. You can state it as a rule with a tell: a concrete signal a reviewer or a grep can catch next time (
kind ===, a hardcoded hex color,expect(x).toHaveBeenCalled(), a duplicated constant). "Write better code" is not a tell: discard. - Not already covered. Search the current skill. If the rule is already there, do NOT duplicate it; at most tighten the existing wording. Most near-misses land here.
- Portable. It belongs in the cross-repo standard. Repo-specific facts (a path, a command, a token name) go to that repo's agent instructions (
CLAUDE.md,AGENTS.md,.cursor/rules, or the equivalent), not here. Anything a formatter or linter already enforces is not a lesson: it is a gate config (Mechanical gates).
Discard (do not touch the skill) when: it is a one-off domain bug; a style preference the formatter or linter owns; project-specific (it goes to the repo's agent instructions); already covered; or a nit with no general signal. Discarding is the common, correct outcome: most comments are SPECIFIC.
Output a short table: each comment, GENERIC or SPECIFIC, and for a generic one the target section and the tell.
Step 2: Fold the GENERIC lessons in, minimally
For each GENERIC lesson:
- Prefer sharpening an existing bullet over adding a new one. Add the missing tell or anti-pattern to the rule that already exists. Add a new bullet only when no existing rule covers it. Add a new section only when a whole dimension is missing (rare; justify it).
- Write in the skill's voice: an imperative rule, the tell, and optionally the anti-pattern. Terse. Match the density of the text around it. No preamble, no hedging.
- Keep it tight. The bar for this skill is high signal, no bloat. If your addition makes a section longer without making it sharper, cut it back. One vivid tell beats three paragraphs.
- One section per lesson. If a comment spans two sections, split it.
Step 3: Show the diff, apply, log
- Show the exact before and after edit to
code-check/SKILL.mdfor each accepted lesson. Keep edits surgical (a tightened sentence, an added tell); never rewrite a section. - Apply the edits.
- Log provenance in
code-check/LEDGER.md, one line per accepted lesson:YYYY-MM-DD · <repo> PR #<n> · <section> · <the one-line lesson>. The ledger is the audit trail: it lets anyone see why a rule exists and revert a bad addition. Use absolute dates. Keep only the one-liner in the ledger; the content lives in the skill. - Report: N comments reviewed, M generic (folded into which sections), N minus M specific (discarded, with why in one phrase). If M is 0, say so plainly. A review with no general lesson is normal and honest; do not invent one to justify an edit.
Guardrails
- The skill must not only grow. When you add, look for a rule the new tell makes redundant, and merge them. Net size stays roughly flat; density rises.
- Never weaken a rule to match a shipped mistake. Lessons raise the bar, never lower it (as with coverage: raise, never lower).
- A shipped regression changes the test class, not just the code. If the lesson is "the suite stayed green while the feature broke", the edit is a sharper Testing tell about the missing corner or seam. Then search the codebase for tests of the same shape.
- Don't over-fit to one loud reviewer. One bot's style preference is not a standard. The "would it recur elsewhere?" test must pass on its own merits.
Scanned by
NVIDIA SkillSpector
Automated static analysis (no LLM) · v2.12.0 · 2026-11-20
Partial analysis: some checks could not run on this component.
0 observations
SkillSpector looks for patterns such as shell commands, credential access, network calls or unpinned dependencies. An observation is something worth reading in the files before you install, not a confirmed vulnerability. Components that run code locally, like hooks, MCP servers and security tooling, naturally show more of them. Open one to see the lines it points at.
No observations.
Learn more about NVIDIA SkillSpector or how components are reviewed.
Agent compatibility
Checked by NVIDIA SkillEvaluator Harness Check · v1.0.1 · 2026-11-20
How each agent reads this skill once it is installed, from the agents' documented loading rules. No agent was run. A note means the agent loads the skill but changes part of its text, or drops it.
Claude Code
loadsCodex
loadsCopilot CLI
loadsOpenCode
loadsHermes
loadsOpenClaw
loads
Every agent loads this skill as written. No notes.
Learn more about NVIDIA SkillEvaluator.