Create Claude skills following best practices: structure, naming, descriptions, progressive disclosure, testing, and converting sources into skills. Use for authoring or improving skills, or encoding a book, guide, codebase, PR history, or visual reference corpus into skills. Triggers: create a skill, SKILL.md, taste.md, convert this book/repo/images to a skill.
Instrucciones de origen · Vista previa de solo lectura
name
creating-skills
description
Create Claude skills following best practices: structure, naming, descriptions, progressive disclosure, testing, and converting sources into skills. Use for authoring or improving skills, or encoding a book, guide, codebase, PR history, or visual reference corpus into skills. Triggers: create a skill, SKILL.md, taste.md, convert this book/repo/images to a skill.
Skills effectiveness depends on the underlying model.
Model
Consider
Haiku
Does the skill provide enough guidance?
Sonnet
Is the skill clear and efficient?
Opus
Does the skill avoid over-explaining?
SKILL.md Structure
---
name: skill-name (lowercase, hyphens, max 64 chars)
description: What it does + when to use it (max 1024 chars)
---
[Body content - under 500 lines]
⚠️ House divergence — this corpus uses noun phrases, not gerunds (House Style rule 4 below): "hardware-product-design", not "designing-hardware". The gerund form below is the official suggestion; both are sanctioned, but match the corpus you're in.
Official suggestion: gerund form (verb + -ing):
Good
Avoid
processing-pdfs
helper
analyzing-spreadsheets
utils
managing-databases
documents
Writing Descriptions
Always write in third person. The description is injected into the system prompt.
Template (official minimal shape — in this corpus, use the fuller House Style rule 1 shape: what + when + boundary clause + triggers, ~350–450 chars):
Rules layered on top of the official guidance, derived from a ~40-skill corpus with a routing eval:
Descriptions: ~350–450 chars, never the 1024 cap. The cap is an API limit, not a budget; all descriptions preload into every session (~40 × 450 ≈ 4.5k tokens here). Shape: What it does (verb-led). Use when <3–5 situations>. <Boundary clause>. Based on <credit>. Triggers: <6–8 query-like terms>. The validator warns >700.
Boundary clauses are sacred. Dense corpora have near-neighbors (8 Rails skills, 5 SwiftUI skills, 6 motion/polish skills). Clauses like "DEFAULT school; escalate on named pains" or "Distinct from X's perf audit" exist because of real routing misses — compress, never drop. The official one-liner examples assume no neighbors; the shape generalizes, the length doesn't.
The routing eval is the gate. After any description change, new skill, or consolidation, run routing probes: give a fresh agent the descriptions only and a set of realistic prompts, and check each routes to the intended skill. Add 1–2 probes (plus a control that must NOT route to the new skill) with every new skill. In this repo, the probe method and fixtures live in docs/ (see the routing-probes sections of docs/skill-library-ops.md; scripts/check_vercel_routing_probes.py verifies the recorded probe packet). This is the docs' "build evaluations first" applied to discovery.
Naming: noun phrases by task shape, never by industry (house divergence from the official gerund suggestion above — both are sanctioned; consistency within the corpus is what matters). "hardware-product-design", not "designing-hardware" or "fintech-skills".
Body budget: <5k tokens (Agent Skills spec recommendation), 500-line hard max. On compaction only the first 5,000 tokens of an invoked skill are re-attached (25k shared across skills) — front-load the load-bearing rules; push long quotes, walkthroughs, and per-source detail to references/. Reference files >100 lines get a ## Contents block (partial reads see full scope).
Knowledge-skill anatomy (talk/article-derived skills): Sources block with titles+speakers → task-shape framing sentence → principle sections with verified, cited quotes (short by default; larger chunks are fine when the passage itself is the load-bearing artifact — always attribute: speaker, work, video ID/URL) → Checklist → "Relationship to other skills" (bidirectional) → staleness note separating durable doctrine from fast-decaying specifics.
Gotchas are the highest-signal content (Anthropic's internal finding). Populate from observed failures, not anticipation — the corpus analog is the verified-quote + failure-driven-rule discipline (e.g. source-sweep was born from a real sampling failure).
Don't restate what Claude knows. A skill earns its place with non-obvious, source-verified material: numbers, named techniques, counter-intuitive rules, decision boundaries. "One skill per task shape; the best skills fit cleanly into one category."
Corpus governance: extend existing > new skill > skip; new skills need ≥2–3 independent sources and a coherent task shape (seed-and-wait in a ledger until then); consolidate clusters under umbrella skills (encapsulation) when descriptions start competing; README.md table = human index; validate_skills.py after every change.
Field triggers for making skills (Dive Club, 2026): build skills from compiled sources when docs are agent-unreadable — Kris Puckett hit this on Liquid Glass: "Apple's developer library… is all in JavaScript. LLMs don't read JavaScript very well," so he copy-pasted excerpts to a research agent — "let's compile it all into one skill" (nPyxVMd1LIA). The trigger is repeat-pasting the same material into prompts. Personal explainers written as you learn are proto-skills — "creating explainers as I go with Claude" (Flora Guo, mdV8APhz2j4). And maintain rules by correction, not anticipation — "if I get an output that I don't like, I'm like, yo, never do that again. This is what I want instead" (Tommy Geoco, OYNoy468kS8) — the lightweight everyday form of rule 7's failure-driven discipline.
Batch quote verification (this corpus): verify quotes with python3 scripts/verify_quotes.py SOURCE... <<'EOF' (one quote per line; or --manifest quotes.json) — one call for all quotes instead of per-quote greps. It encodes the accumulated normalization rules (curly quotes, OCR mid-word spaces, transcript stutter, ellipsis-split fragments, accent folding). A MISS means do not ship the quote: re-read the source region and fix the wording or drop it. Pipeline rule: when a screening agent has already verified quotes, it should emit a quotes.json manifest; downstream fold agents copy quotes from the manifest verbatim (never retype from a prompt) and exhaustively verify only quotes they add themselves, spot-checking ~20% of manifest quotes.
Citation economics (this corpus): inline citations are short name-tags only — "(Rutter)", "(Klein, ch. 9)", speaker name on a quote — because the NAME does routing work between competing schools and survives editing (numbered [1] refs rot under parallel folds). The bibliographic apparatus (full titles, years, URLs, video IDs, session numbers) lives in references/sources.md per skill, with SKILL.md carrying a single pointer line. Sources blocks at the top of a body spend the compaction-surviving budget on provenance — keep the hot zone for rules. Operational notes (⚠️ supersession, caption-garble keys) stay in SKILL.md; pure bibliography moves.
Taxonomy & graph rules (docs/taxonomy-2026-06.md): top-level = a platform, an activity, or a cross-domain tool; topics nest as reference files under their platform/activity cluster. Nesting never severs access — every inbound edge from another skill becomes an explicit cluster (member reference) cross-ref. Any consolidation requires: corpus-wide edge remap, an independent grep audit (zero dangling member names), and before/after routing probes that gate the change (revert if AFTER < BEFORE). Absorbed members keep a *Scope:* line; cluster descriptions pool member trigger vocabulary (≤700 chars); user-invocable commands stay top-level.
Write tight from the start (compaction-pilot finding, 2026-06: bodies compress only ~10% after the fact without content cuts — so author lean instead): clipped fragments over full sentences for example lists; arrow chains for procedures (audit → inventory → group); quote-intro framing cut to "Speaker, context:"; no connective glosses ("this is the hinge..."); cross-ref bullets = distinction + routing rule only. Never compress quotes, rules, or distinct examples — density lives in the framing, not the content.
Agent-standard architecture (Vercel product-design pattern, 2026): when a skill teaches agents repeated product/design judgment, split the system into trigger/routing, source-grounded guidance, exemplars, coverage gaps, deterministic checks, and a human update loop. Start from repeated review decisions; write scope, rationale, evidence, exceptions, bad/good example, and approver before promoting a candidate into a reusable rule. Use lint/scripts only when the failure is mechanically identifiable and has a concrete fix; keep judgment in prose. Test retrieval separately from application.
Register skills where work is created, not only where it's requested (Pocock/Ness wayfinder pattern, 2026): description-matching at request time is one routing surface; a planning skill that charts work into tickets is another. Its ticket-type definitions and plan notes can name the skill that resolves each ticket — Will Ness wired his new frontend-prototyping skill into Pocock's /wayfinder planner with one routing sentence, so every future plan containing novel frontend work generates a ticket that invokes it ("this is a very powerful pattern for planning work"). When authoring a skill meant to fire at a specific workflow moment, add a routing clause to the orchestrator/planner skill that creates that moment's work — the workflow-mechanics side lives in working-with-ai (agentic-coding).
