On a database migration task, the same skill scored 90% with a smaller model and the model hit 100 out of 100 without any skill at all. That contrast is the whole argument behind Claude Code skills: they are not free wins, they are targeted instructions whose value you must measure per model before shipping them.
What Claude Code skills are and where they live
Claude Code skills are folders of reusable instructions that Anthropic terminal coding agent loads on demand to perform a specific task. In a project that uses Claude Code, you keep them under a skills directory, one folder per skill, with a single SKILL.md file inside. In other agent setups the same directory may be called agents, but the structure stays the same.
This walkthrough comes from a video published in August 2026 by the Full Cycle channel, presented by Gustavo Dev Doido, who teaches in the Crazystack Typescript program and its Bootcamp do Dev Doido. The demo shows two custom skills and one official skill from Anthropic, plus the evaluation workflow used to decide whether each skill is worth keeping.
The speaker's position is direct: without structured instruction files, workflow-driven processes are hard to run reliably today, because the model needs stable context about how your team wants the work done.
What goes inside a SKILL.md file
A SKILL.md file has two parts: a YAML frontmatter block and a body. The frontmatter carries the name and the description, which is what the agent reads to decide when to activate the skill. The body is the actual instruction, written in plain language, that tells the model what to do and how to deliver the result.
A minimal skill can be nothing more than that one file. More complete skills add two things alongside it:
- References — extra context files, such as a catalog of SQL performance antipatterns, that the skill can point to when the task needs deeper detail.
- Deterministic scripts — small programs the agent runs instead of improvising, which makes the outcome repeatable.
The official skill-creator skill from Anthropic follows the same layout but is richer: it ships a central SKILL.md, scripts, references, and supporting files that let it run evaluations on your other skills.
Three example skills from the demo project
The video shows three skills side by side, which is a useful sample of the range a real project accumulates.
SQL performance antipatterns
The first skill is database-focused. Its body contains the instruction, and its reference folder lists items to watch for when writing SQL, such as known performance antipatterns. The reference exists so the model does not have to guess what 'good' looks like.
Database migration
The second skill also targets databases but automates a process: migrating from one database to another. It demonstrates that a skill can encode a multi-step workflow, not just a style rule.
Skill creator
The third is Anthropic skill-creator. It is responsible for testing your skills through an evaluation process, and its extra folders exist to support that job. You did not write it, and you do not need to modify it; you just invoke it.
How the description evaluation loop works
The evaluation loop answers one question: does the skill trigger when it should? In the demo, the speaker deliberately broke his skill by truncating half of its description, then asked skill-creator to run the evaluation focused on that description.
Skill-creator executed its run loop file and iterated: it generated candidate descriptions, tested whether the skill fired correctly for each one, and analyzed the results. At the end it reported the best description found, and the broken skill started behaving correctly again.
Everything the loop produces is stored in a workspace directory. There you find the query list used during the loop, each labeled with a should-trigger or should-not-trigger parameter, plus the logs of every iteration. That trail is what makes the optimization auditable rather than magical.
Which model benefits most from a skill
The second evaluation compares model quality on the database migration task, run twice in parallel: once with the Haiku model and once with Opus, both Claude models documented by Anthropic. The speaker reported these numbers from his own run, so treat them as one first-hand demonstration rather than a benchmark.
| Configuration | Result reported in demo | Reading |
|---|---|---|
| Opus, without the skill | 100 out of 100 | The skill adds nothing on the strongest model |
| Opus, with the skill | 100 out of 100 | No measurable gain |
| Smaller model, with the skill | about 90% | Skill justifies its existence |
| Smaller model, without the skill | about 50% | Frequent qualitative failures across several items |
The 40-point delta between with-skill and without-skill runs on the smaller model is the core lesson. Skills matter most where the base model is weakest. On a frontier model that already solves the task alone, the skill is redundant, and the evaluation tells you that before you ship dead configuration to your team.
How to read the evaluation report
Beyond the chat output, the evaluation produces a browser report with an executive summary. For each prompt in the test set, you see the answer the model gave, the grades, and the graded items themselves.
In the demo, the first prompt passed five out of five, while a later one passed two and failed three. Each failure shows why it failed, which is the practical payoff: you take the failing item, adjust the skill, and re-run. The speaker's point is that this loop lets you improve a skill progressively instead of guessing.
The report also consolidates iterations, showing which run used which model and whether the skill was present, so you can trace every score back to its exact configuration.
Making skills work across a team's different agents
A workflow only pays off if it behaves similarly for everyone on the team, and teams do not standardize on one tool. The speaker names the problem concretely: one developer prefers Claude Code, another uses Codex, a third uses OpenCode.
Because skills follow a portable folder convention rather than being locked to one client, the same instruction set can be carried across tools, and the evaluation process gives you a way to verify that the experience stays roughly comparable everywhere. Without that check, a workflow tuned for one agent silently degrades on another.
Frequently asked questions
- What is a Claude Code skill in one sentence? It is a folder under your project's skills directory containing a SKILL.md file with a name, a description, and instruction text, plus optional references and scripts.
- Where does the description matter most? In the frontmatter: the agent reads the description to decide whether to activate the skill for the current request, which is why a truncated description broke the skill in the demo.
- What is skill-creator used for? It is Anthropic official skill that runs evaluations on your skills, iterating over candidate descriptions and prompt sets to find and validate the best configuration.
- Do skills help every model equally? No. In the demo the strongest model scored 100 out of 100 with or without the skill, while a smaller model jumped from roughly 50% to 90% with it.
- How do I know if a skill is failing? Run the evaluation report, inspect each graded prompt, and read why the failed items failed; the report itemizes scores per configuration.
Turn your own walkthroughs into written guides
The video's central lesson is that reusable knowledge, whether a skill or an evaluation method, only pays off when it is written down and measured. The same is true of the explanations developers give on camera: a twelve-minute screen recording full of folder structures and score deltas deserves to exist as an article people can search and cite.
If you have that kind of knowledge inside your own YouTube videos, Skala Blog converts them: paste the URL, the video is transcribed, and a structured article is generated for you to review and publish.
Fork this article
Start a new branch from the same video, shaped your way. You keep the credit; the original keeps the attribution.
A fork in another language is filed as a translation of this article, so the two pages point at each other. You can unlink it later from the editor.
0/240
You are creating
- Format
- For
- Language
- Source
- Your angle
You will be asked to sign in before it is generated.
Buy credits