Create a skill that the agent chooses
A description that triggers, a short body, and progressive disclosure: each file enters the context only at the step that needs it.
The problem
You wrote the skill. The agent sometimes uses it, sometimes ignores it.
A poorly written skill costs context and delivers a different result on every run. The problems below are the most common, and each course in this training program addresses one of them.
- The description doesn't contain the words you type, and the agent doesn't choose the skill.
- The SKILL.md became an 800-line manual, and the rules at the end are never read.
- A skill written for the old model over-explains and makes results worse on the new model.
- The rule that must not be broken is in capital letters in the text, and it still fails sometimes.
- The same skill works in Claude Code and breaks in Codex.
After the training
By the end of the 7 courses, you can:
Learn from real skills
Twelve skills dissected from the inside, from an itinerary generator to an SEO auditor and a RAG architect.
Write with engineering rigor
Grilling, CONTEXT.md, TDD, and the failure modes of Matt Pocock's skills, the same ones the validator uses.
Cut what's left over, by measuring
Ablation audit on CLAUDE.md and on skills: an A/B/C plan, cutting only what the test proves to be dead weight.
Take the skill to Codex
Verifiable skills in Codex, freedom calibrated to risk, and verification with evidence.
One skill for both agents
CLAUDE.md and AGENTS.md side by side, the polyskill CLI, and two agents working on the same project.
7 courses in order · 93 lessons and modules
Seven courses, in the order the area recommends.
Build your first one, learn to write well and audit, then take it to Codex and to the team. If you already have skills, you can enter at course 3 and run the validator before rewriting.
- 1
First skillPT · EN · ES
Skills v6
For people who don't code: what a skill is, your first skill, and how to make it good, by testing and fixing it in the Claude app, with no terminal.
8 lessonsOpen course 1 - 2
AnatomyPT · EN · ES
Claude Skills in Practice
The anatomy of SKILL.md, progressive disclosure, descriptions that trigger, and 12 real skills dissected, from a sales funnel to multi-agent memory.
17 modulesOpen course 2 - 3
Criteria
Skills For Real Engineers
Matt Pocock's skills: grilling, ubiquitous language with CONTEXT.md and ADRs, TDD, composition, and skill customization.
12 modulesOpen course 3 - 4
AuditPT · EN · ES
Ablation Audit
Why configuration ages, the taxonomy of each line, the audit-ablacao skill running, and the A/B/C plan for cutting without losing quality.
8 modulesOpen course 4 - 5
ChassisPT · EN · ES
HARNESS (Matt Pocock)
Prompts, skills, environment, and code as the chassis you control: a skill's context cost, grill-me, chaining skills, and AFK setup with review.
30 modulesOpen course 5 - 6
CodexPT · EN · ES
Agent Skills in Codex
A skill with a clear job, calibrated freedom, verification with evidence, visual review, and a comparison of quality, cost, and effort.
8 modulesOpen course 6 - 7
Two agents
PolySkills
Claude Code and Codex side by side: CLAUDE.md and AGENTS.md, skills in both, assisted migration, the polyskill CLI, and two agents on one project.
10 modulesOpen course 7
Who it's for
Made for you if
- You keep repeating the same instructions to the agent and want it to do things your way without being asked again.
- You already have skills and don't know which ones the agent uses, which ones it ignores, and which ones got worse with the newer models.
- You use Claude Code and Codex and want a skill that works in both.
- You want to build a skill library for your team, with validation before publishing.
Not for you yet if
- Do you want to change the Claude Code screen with panels and guards? That's a mod, in the MODS area.
- You're looking for a standalone prompt for a single task, without repeating the process.
Included in the training
Alongside the courses, ready-made skills you can open up and look inside.
The area brings together 43 INEMA projects that are already skills. Install them, run the validator, and compare them with the rules the courses teach.
Projects and kits
Why this training
What changes compared to copying ready-made skills.
Updated for the 5.5 models
The area brings together Anthropic's skills guide, the Claude Code docs, and the prompting guide for Fable 5 and Opus 5.5: less manual, more verification.
Open validator
auditar-skills checks 46 rules in one second, without installing anything, and works with skills in Portuguese and Spanish.
A criterion you can measure
A skill works when the agent takes the same path on every run. The courses teach you to test with and without each line.
Claude Code and Codex
Courses 6 and 7 take the same skill to Codex and to both agents in the same project.
Questions
Before you start.
Do I need to know how to code?
Not to get started. Course 1 creates and tests a skill in the Claude app, with no terminal. The following courses bring hands-on practice in Claude Code or Codex.
Do I need to follow the order?
The suggested order is 1 → 7. If you already have skills, you can start with course 3 or 4 and run the validator on your own library.
Does it work in Codex?
Yes. Skills are folders with SKILL.md, which Claude Code, Codex, and other hosts read. Courses 6 and 7 cover Codex.
How long does it take?
There are 93 lessons and modules across 7 courses. The shortest have 8; HARNESS, the longest, has 30.
What is the best first step?
Run the auditar-skills validator on the skills you already have. It takes a minute, uses no tokens, and shows where the new rules apply.
Your next skill can make the agent work your way without you repeating the instruction.
Start with course 1. In 8 lessons you create, test, and fix your first skill, with no terminal.