Agent skills

The wizard's last act is to hand the project to a coding agent, so the app has to arrive with notes and not only with code. Everything on this page follows from that: the agent that continues the work needs what we know about the API's edges, and it needs it in the place its own CLI looks.

One skill per module

Each module carries a skill/ directory in this repository, and the wizard copies it into your project as chatfuel-<id> — chatfuel-deals, chatfuel-livechat, one per module you picked. Each one holds that module's structure, its GraphQL operations and the traps in the API behind it.

Beside the copy goes a small sidecar recording the module, the wizard version and the timestamp. That is what lets a later run tell its own installs apart from a directory you wrote yourself: it names the version and the date and asks before replacing, and a directory it does not recognise gets a blunter question.

Two layouts, because neither CLI reads the other's

Neither reads the other's directory, so this is a choice the wizard has to make rather than hedge. Copying both layouts would leave two trees to drift apart instead of one to maintain — so the skills are written once, and moved if the agent you settle on at the handoff is not the one they were laid out for.

The difference goes one level deeper than the paths. Claude Code loads its skills directory by itself, so naming the skills is enough. Codex's discovery depends on its version and the user's config, so its instructions file spells out the full path to each SKILL.md.

Skills can also go to your home directory instead of the project, in the same layout — useful when you will scaffold more than one app and want one copy of the notes.

The handoff note, and the checklist it lands in

Two files come out of the run, and they do different jobs.

CLAUDE.md or AGENTS.md is the standing brief: the routes this app has, the skills to read before touching the API, which environment variables are secrets that must never be printed, and how to resume the guided setup later. It is short on purpose — it is read on every session start.

The setup checklist is the one-time script for the first session: how to greet you, a tour of the modules that were installed, the deep links to what was seeded, and the instruction to start the dev server rather than ask you to. Each module's handoff.md is inlined into it verbatim, under a "Module guides" heading, which is where the module-specific first tasks and gotchas live.

Claude Code reads that checklist as a slash command; Codex reads it as a skill, with the frontmatter that makes the same text valid in both places. When the wizard launches the agent for you, all it types is the instruction to run it.

In --embed mode the instructions file belongs to a project that already existed, so the wizard writes a marked section into it and rewrites that section on later runs instead of overwriting the file.

What a skill is actually for

The concrete case: module code is plain React and a typed GraphQL client, and the operations each module uses live in its skill's examples, validated against the bundled schema. So "add a column to the deals table" is a sentence rather than a file hunt — the agent already has the operation, the schema and the note about what that field does when it is null.

That is also why the skills are per module rather than one big document. You installed four modules, so your agent reads four sets of notes and not thirteen.

This site does the same thing for the same reason. It serves itself as a skill at /SKILL.md, and every documentation page as Markdown — /docs/concepts/state.md is this section's neighbour in plain text — with the whole list at /docs/llms.txt.

What it costs

A skill is a copy, like everything else the wizard writes. It reflects what we knew on the day you ran it, and it does not update itself — re-running the wizard is what replaces one, after it has asked. If you edit a skill to match your own changes, note that a later run will offer to overwrite it.

Your first change ends with the version of this you can try in one line.