Build a skill

An Agent Skill that turns the conversation you are in into a new skill with evals. Use it the moment the agent has produced what you wanted after one or more corrections: it finds where the story started, captures every round, writes the skill from those corrections only, scaffolds a repository with eval suites, replays your first prompt with the skill installed, and proves the skill with runs with and without it.

It works in Claude Code and Codex and is written to be agent-independent: the folder follows the Agent Skills specification, and the only agent-specific parts (where transcripts live, how to install) are listed in transcripts.md and replay-and-prove.md. Another agent can ask you for what it cannot find.

Install

From the Tessl registry, into the agents configured in your project:

tessl install martinfrancois/build-a-skill

Or from a clone; a symlink keeps the installed skill in step with this repository.

git clone https://github.com/martinfrancois/build-a-skill.git
cd build-a-skill
# Claude Code
mkdir -p ~/.claude/skills && ln -s "$PWD/skills/build-a-skill" ~/.claude/skills/build-a-skill
# Codex
mkdir -p ~/.agents/skills && ln -s "$PWD/skills/build-a-skill" ~/.agents/skills/build-a-skill

Use

In the conversation where the agent finally got it right:

/build-a-skill          # Claude Code
$build-a-skill          # Codex

or in plain words: "turn these corrections into a skill". The skill then:

1. finds the first prompt that produced the almost-right output, from the conversation or the session transcript on disk, and asks you if it cannot tell;
2. captures each round: code before, prompt, output, your correction, the version that stayed, and why;
3. proposes a name and a directory for the new repository;
4. copies the template there and writes SKILL.md, its examples as "not like this, like this, why" pairs, and at least three eval scenarios, one or more per round;
5. replays your first prompt at the commit before it, with the new skill installed, and compares the result with the version that stayed;
6. runs the quality review and the evals until the review and every with-skill scenario score 100 and at least three scenarios sit in the main suite, with a stop rule of three runs per scenario, and installs the skill. The Tessl registry gives a full aggregate score only to a main suite of three or more.

The template's README holds the same steps for doing them by hand.

Credits

Hosted evals use Tessl credits: as of October 2026, Free includes 1,000 per calendar month and a paired scenario costs 20; since November 2026 a quality review costs 10. The skill checks the balance with tessl org usage --json before every hosted run, the review included, spends credits in order (targeted scenarios, then suites), and stops at the first stage that is not clean. When the month's credits are used up it runs the scenarios locally against the same model for a cent or two each instead.

License

MIT, copyright 2026 François Martin.