Skills: the guide

The 57 Éclair skills, in 8 families, help your AI assistant add a feature to your Éclair 2.0 site without breaking what works. Install them in your project, start the agent with the supplied prompt, and activate nothing before the 7 quality gates.

This guide details each step and explains how to write your own skills. Before you start, make a backup as explained in the Updates, backups and moving guide.

What do you need before you start?

• An Éclair 2.0 site and, ideally, a working copy of the code (the official repository cloned), so the agent can run the tests.
• Python 3 to install and validate the skills, PHP 8.1 or later for Éclair’s tools.
• An assistant that reads skills: Claude Code or Codex, for example.
• A recent backup: Security › Download a backup.

How do you install the skills in your project?

1. Download the library from GitHub (Code › Download ZIP, or git clone).
2. In its folder, check it is intact with python3 scripts/validate_skills.py: the message should report 57 valid skills.
3. List families and skills with python3 scripts/install_skills.py --list.
4. Install what you need, by family (--category design) or by name (--skill eclair-seo-audit), with --project /path/to/my-site.
5. For Claude Code, copy the installed folders from .agents/skills/ to .claude/skills/.

python3 scripts/validate_skills.py
python3 scripts/install_skills.py --list
python3 scripts/install_skills.py --project /path/to/my-site --category foundation
python3 scripts/install_skills.py --project /path/to/my-site --skill eclair-wordpress-migration
python3 scripts/install_skills.py --project /path/to/my-site --all

How do you start the agent with the right prompt?

Open your assistant in the project folder and give it a clear starter prompt. This one ships with the library:

Analyse this Éclair CMS project starting with eclair/index.php and docs/SKILLS-GUIDE.md. Use the skills eclair-architecture, eclair-extension-builder, eclair-security-audit and eclair-test-suite. Describe the requested feature in terms of the native eclair-extension/1 capabilities. Create the extension files under site/extensions/<id>/ without modifying the core. Before proposing activation, run php tools/validate.php <extension> --json, php tests/run.php and php tools/measure.php --check. If a capability is missing, stop that part and create an RFC and tests; do not invent an API.

Then add your request in one or two sentences, with the constraints that matter to you: audience, languages, volume, personal data.

What checks before activating a feature?

The library defines seven quality gates. A feature is only activated once all of them are passed.

What to do

• Describe the business need, not the technique: the agent picks the skills.
• Let the agent read the site’s capabilities before writing code.
• Test on a copy of the site, never directly in production.
• Keep a record of decisions: the agent writes a short architecture note.
• Plan the rollback: deactivating a feature keeps its data.

What not to do

• Modify app/, assets/, index.php or the updater: an update would overwrite everything.
• Activate a feature that hasn’t passed the validator and the tests.
• Put a cart, a member area or a forum into production while the core doesn’t provide the contract.
• Believe an SEO, GEO or Lighthouse score is guaranteed: they are measurements to repeat.
• Give the agent secrets in plain text: they belong in data/ext/<id>/secrets.php.

How do you create your own skill?

skills/eclair-my-trade/
├── SKILL.md
└── references/
    ├── acceptance.md
    └── native-contract.md

1. Create a skills/eclair-<name>/ folder with a SKILL.md file.
2. At the top of the file, set name and a description that says when to use it: this is what the agent reads to decide.
3. Write the mandatory diagnosis, the procedure, the deliverables, the security invariants and the acceptance criteria.
4. Add your useful documents in references/: a skill must be self-contained.
5. Declare it in catalog.json, then run python3 scripts/validate_skills.py and python3 scripts/check_release.py --write.
6. Test it on a real project, then propose it to the library with a pull request.

What if something gets stuck?

The validator rejects the package: Read the JSON message: it names the file or manifest field at fault. Fix it, then run the validation again.

The agent wants to modify the core: Refuse and restate the rule: everything must live in site/. If a capability is missing, ask for an RFC and a blocking test.

A capability is missing (customer accounts, payment…): That’s expected for some families: keep the design and the tests, and wait for the core to open the contract.

See the library on GitHub

The 57 skills, family by family