🧰 What is a meta-skill
A meta-skill is a skill whose job is to help you work with other skills: discover, create, test, optimize, and package. While an ordinary skill handles the end user’s task (editing an xlsx, building a dashboard), a meta-skill works one level up—it’s the workshop where skills are born. For skill creators, mastering three meta-skills changes the game.
The three that matter
skill-creator — anthropics · 246k installs
The canonical workshop: captures intent, drafts, runs evals, iterates, and optimizes the description.
find-skills — vercel-labs · 1.802.925 installs
The most installed skill in the entire catalog. It finds existing skills before you create one from scratch.
scaffolding / templates — a pattern, not a single package
Generates the initial skeleton (SKILL.md + folders) so you don't start with a blank page.
💡 Why start with the tools
In the catalog of 39.366 skills, 60,2% have fewer than 100 installs. Many people create skills without ever checking whether something better already exists or using the eval loop. Meta-skills solve exactly that: discover before creating, create methodically, and validate with data.
⭐ skill-creator — the canonical workshop
O skill-creator from Anthropic is the reference meta-skill. It doesn't just write a SKILL.md: it orchestrates the entire cycle — draft → eval → iterate — and includes a separate description optimizer. It's the foundation for everything you saw in modules 4.1 and 4.2.
What it does
Capture intent (4 questions), write the draft, generate test prompts, run with-skill vs. baseline in subagents, aggregate the benchmark, and run the description optimization loop.
When to use
Whenever you’re creating a new skill, improving an existing one, running evals, measuring performance with variance, or optimizing triggering. It’s the starting point—don’t improvise a parallel process.
How it fits into the workflow
It’s the central axis. find-skills comes first (discovery), and scaffolding happens inside it (generating the foundation). skill-creator ties the scripts together: aggregate_benchmark, generate_review.py, run_loop.py, package_skill.py.
skill-creator frontmatter (description is the trigger):
name: skill-creator description: Create new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, edit, or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy.
🎯 Pay attention to the description
It lists concrete contexts ("create from scratch", "run evals", "benchmark", "optimize description") instead of a vague phrase. The meta-skill practices the advice it gives: the description should say what it does AND when to use it, and be a little pushy.
🔭 find-skills — discover before creating
find-skills (vercel-labs) is the most-installed skill in the catalog: 1.802.925 installs. Your job is to discover skills relevant to a task — which prevents the costliest mistake a creator can make: spending hours writing something that already exists in a better form.
✗ Create blindly
- ✗Writes from scratch without checking the catalog
- ✗Reinvent a skill with 200k installs and years of iteration
- ✗Becomes one more of the 60.2% with <100 installs
✓ Discover first
- ✓Runs find-skills with the task description
- ✓Finds what already exists and uses it, extends it, or draws inspiration from it
- ✓Only create from scratch when the niche is truly empty
concept of the return from a skills search:
{
"query": "transformar planilha xlsx em relatório",
"matches": [
{ "name": "supabase-postgres-best-practices",
"installs": 203000, "source": "supabase" },
{ "name": "frontend-design",
"installs": 488299, "source": "anthropics" }
],
"decision": "estender existente | criar nova"
}
💡 Discovery is step zero
find-skills comes BEFORE skill-creator. The healthy flow is: describe the task → find-skills → decide (use / extend / create) → only then open skill-creator. Skip this step and you risk duplicating work another team has already done better.
🏗️ Scaffolding and templates
Scaffolding is the standard for generate the initial skeleton of the skill instead of typing everything by hand: the SKILL.md with frontmatter, the folders scripts/, references/, assets/ and template files. It’s not a single famous package—it’s a technique that skill-creator itself applies internally when it starts a draft.
skeleton generated by the scaffolding:
minha-skill/
├── SKILL.md # frontmatter (name + description) + corpo
├── scripts/ # código determinístico/repetitivo
├── references/ # docs carregadas sob demanda
├── assets/ # templates, ícones, fontes
└── evals/
└── evals.json # test prompts (sem assertions ainda)
✓ Good use of a template
- ✓Generates the standard structure and the right empty files
- ✓Pre-fills the frontmatter with name + description placeholder
- ✓Lets you focus on the content, not the plumbing
✗ Template becoming a crutch
- ✗Creates
scripts/ereferences/empty "just in case" - ✗Leaves the template boilerplate in the final SKILL.md
- ✗Forget to cut what isn't pulling its weight (violates "keep it lean")
💡 Create only the folders you’ll use
The skill-creator’s advice is “don’t create everything upfront — create directories as needed.” Scaffolding speeds up the start, but empty folders confuse the model and the reader. Generate the skeleton, then prune the structure.
🔗 Where Each One Fits in Your Workflow
The three meta-skills don’t compete — they connect in sequence. The sequence that avoids rework and creates lasting skills is simple and has a specific order.
🔭 find-skills — discover
Before anything else: does something already exist? Use, extend, or create? A decision based on what the catalog of 39.366 skills offers.
🏗️ scaffolding — generate the base
Initial skeleton inside skill-creator: SKILL.md + the right folders. A non-empty page.
⭐ skill-creator — create and iterate
Draft → test prompts → evals → benchmark → iteration loop until it meets the requirements.
🎚️ Description optimizer — fine-tune triggering
Still within skill-creator: should-trigger / should-not, choose best_description by test score.
The golden rule
Discover before you create. Generate the foundation before iterating. Iterate with data before packaging. Each meta-skill covers one stage — using them out of order is where skills with <100 installs are born.
📋 Quick reference table: which meta-skill to use
A practical summary to put on the wall. Given a situation, which tool should you reach for?
situation → meta-skill:
"acho que preciso de uma skill pra X" → find-skills "o catálogo não tem, vou criar" → skill-creator "não sei nem por onde começar o arquivo" → scaffolding "tenho um draft, quero validar" → skill-creator (evals) "a skill não dispara quando devia" → otimizador de description "a skill está pronta, quero distribuir" → package_skill.py
246k · anthropics. The focus: create, evaluate, measure, optimize.
1.8M · vercel-labs. The first step: discover before creating.
Standard. The scaffold: SKILL.md + folders, no blank slate.
💡 Pro tip
Install all three and make them available. Since Claude tends to under-trigger, explicitly say "use find-skills to check" and "use skill-creator to create" the first few times—then it becomes a natural part of the workflow.
✅ Module Summary
Next:
Module 4.4 — 🛠️ How to Create: Complete Walkthrough with Evals — a worked example from start to finish, from intent to evals.json with verifiable assertions.