Menu
What this skill does, leading word first. Then one trigger per branch -- each genuinely distinct situation that should fire it, listed once, no synonym piles. If a trigger collides with a sibling skill, route it: "X is other-skill's job."
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Basé sur la classification professionnelle SOC
| name | skill-name |
| description | What this skill does, leading word first. Then one trigger per branch -- each genuinely distinct situation that should fire it, listed once, no synonym piles. If a trigger collides with a sibling skill, route it: "X is other-skill's job." |
What this skill does and why it exists, in one paragraph. If there is a one-line test for whether the output is good, state it here.
Open by sizing the task. When should the agent say "no" or "not this heavy"? A table mapping task weight to protocol depth beats running the full ceremony on a typo fix.
Named failure modes for diagnosis (reference, not steering). In the process itself, steer positive: state the target behavior instead of prohibiting the failure.
| Reference | When to load |
|---|---|
| references/example.md | Only some runs need it -- say which |
Inline what every run needs; disclose what only some branches reach. The pointer's wording decides whether it gets loaded.
Burnish a repo until the work gleams -- people impute the quality of the work from the quality of its surfaces, so audit, renovate, instrument, and promote open-source repos until real work reads as real in 30 seconds and every claim survives scrutiny. Fires on four branches: polishing or launching a repo ("make this repo world-class", "fix the README", "prepare for launch", "why does this look abandoned"); org-wide housekeeping ("run housekeeping", "which repos are stale", scheduled groundskeeper sweeps); outbound moments ("write the release notes", "announce this", "make people care"); and automation health ("set up a scheduled workflow", "why did nobody notice this broke"). Fixing a single bug is just a fix; product UI design is kiln's job; voice for one standalone document is a writing task, not repo packaging.
Animation implementation protocol for web interfaces: easing, duration, springs, CSS transitions, keyframes, Motion (Framer Motion), WAAPI. Fires on three branches: adding motion to a component ("animate this", "should this animate?"); diagnosing motion that feels wrong ("make this transition smoother", "what easing should I use"); and reviewing pasted animation code. Deciding what should move and why is kiln's job (motion.md); this skill makes it move correctly. Static styling without motion is kiln's too.
Distill a knowledge source -- book, transcript, course, doc set, post series -- into the few agent skills it actually earns. Strict yield: most sources earn 0-3 skills and a digest; every unit must pass four gates (grounded, predictive, earned, behavioral) before it ships. Fires on two branches: the user wants a source turned into skills ("distill this book", "make skills from this course / transcript / doc set"); or they ask whether a source is worth distilling. Capturing a live session's workflow is skillize's job; a skill from scratch is skill-creator's. A summary, review, or reference wiki needs no skill -- write the deliverable directly.
Goal engineering for agentic loops -- a goal is a contract, not a wish. Composes fuzzy intent into a goal brief plus a /goal condition, or sharpens an existing draft against five contract gates (outcome, evidence, boundary, stop rules, pause conditions). Calibrates to task weight: trivial skips, simple gets three lines, complex gets pre-mortem. Fires on three branches: defining a task or setting up /goal or /loop ("I want to build", "help me start", "define the task"); checking whether a goal, spec, or requirement is clear enough ("sharpen this", "is this goal clear"); and recovering after a loop drifted or built the wrong thing. Mid-execution corrections with clear direction don't need a contract -- just make the correction.
HTML artifact protocol for agent deliverables. Decides whether a deliverable beats markdown, picks the artifact shape, builds a self-contained file against a craft baseline, and reviews existing artifacts. Fires on three branches: an explicit ask for an HTML, visual, or interactive deliverable; producing any plan, report, explainer, comparison, deck, or diagram (mermaid / graphviz / d2) whose markdown equivalent would run past ~100 lines; and improving an HTML deliverable that looks like generic AI output. Production frontend code and hand-maintained docs (README, CONTRIBUTING) are not artifacts -- build those directly.
Precision UX design instrument: perceptual principles, composable patterns, parameterized material kits, anti-AI-slop detection. Fires on three branches: designing or restyling any frontend surface (layout, typography, color, spacing, responsive, dark mode, accessibility); auditing or critiquing an existing interface; and rescuing output that looks generic or AI-made ("make it look better", "needs more taste"). Animation design decisions live here (motion.md); implementing the animation is animate-it's job.
