Small, self-contained skills for AI coding agents. Each one exists because something went wrong first.
Skill Claude CodeCodex
Run a second model as a subagent without your repository's ambient context reaching it. A coding agent injects git branch, changed files and recent commit messages into every subprocess prompt, and that background text is scored alongside your actual question -- so a benign request can be declined because of what your…
Skill Claude CodeCodex
Verify every citation before it is published. Fetch each DOI, PMID, arXiv or ISBN from the issuing registry and confirm the title and authors match the claim being made. Use whenever an agent writes, edits or reviews content containing scientific, medical, legal or statistical claims, and whenever a citation is added…
Skill Claude CodeCodex
Stop work evaporating between sessions. Every promise gets one durable record that resurfaces until it is explicitly disposed of; 'parked' and 'killed' are honest endings and a silent drop is the only failure. Use whenever you write 'we should' or 'later' or 'I'll come back to this', whenever a session ends with loose…
Skill Claude CodeCodex
Record a human's decisions so they are not silently re-litigated. Capture the decision in their own words, in one place, the moment it is made; gate re-asking on a lookup so the same question cannot be asked twice. Use whenever a human answers a question that constrains future work, whenever you are about to ask a…
Skill Claude CodeCodex
Never trust a green light that cannot state its denominator. Use whenever you write, review or rely on a check, guard, lint, test sweep or audit; whenever a tool reports 'no problems found', 'all clear' or '0 issues'; and whenever a result seems too easy. A check that finds nothing and a check that checks nothing…
Skill Claude CodeCodex
Decide whether an automated guard should BLOCK or WARN. Block only on deterministic facts; warn and log on judgment. Use whenever you write a hook, pre-commit check, CI gate, lint rule or policy check; whenever a guard fires on something legitimate; and whenever you are tempted to add a bypass flag or to use one.
Skill Claude CodeCodex
A model's self-reported identity is not evidence. Verify which model actually produced an output from the transcript or API metadata, never from what the text claims. Use whenever a second opinion, review or sign-off is attributed to a specific model; whenever a fallback or router may have substituted one; and before…
Skill Claude CodeCodex
Stop tests writing into real data. Withhold production credentials by location so a misdirected test cannot reach production even if it tries, rather than asking test code to behave. Use whenever a test performs a write, whenever you point a suite at an environment, whenever you add a 'safe mode' flag, and whenever…
Skill Claude CodeCodex
Stop spawned processes from popping console windows on Windows. A console child steals keyboard focus mid-keystroke and can cause the user to lose work. Use whenever an agent, script, scheduled task or background service spawns a subprocess on Windows, and whenever a user reports windows flashing or focus being stolen.
Skill Claude CodeCodex
Find WordPress JavaScript errors that appear ONLY for logged-in users, caused by admin-context scripts loading on front-end pages. Use whenever a public page throws a console error that vanishes when you log out, and whenever registering script enqueues -- a logged-out check cannot see this class of failure at all.
Skill Claude CodeCodex
Stop WordPress silently corrupting inline JavaScript. wptexturize converts an ampersand caught between a into an HTML entity, which kills the entire block with no PHP error and no server-side warning. Use whenever emitting or authoring inline JavaScript that reaches WordPress content filters, and whenever inline JS…
At most 3 mods per repository are shown here, and a mod shipped inside a plugin is left to that plugin's page — the rest are on their repository pages: