---
name: bailey
version: 1
description: >-
  Publish an AI-made web page so it gets a live address and remembers what
  visitors do — governed, GDPR-safe persistence with no backend to manage. Use
  whenever the user wants to put something online, give it an address, keep form
  responses, or have a page "remember" a value. Triggers (EN): "publish", "put it
  online", "give it an address", "keep the responses", "collect emails", "it
  should remember". Déclencheurs (FR) : « publie », « mets en ligne », « donne-lui
  une adresse », « garde les réponses », « collecte les emails », « ça doit se
  souvenir ».
---

# Bailey — publish & remember

This skill is deliberately **thin**. It does not carry the how-to: that lives on
the Bailey server and is served fresh on every call, so this file never goes stale.
Your job here is only to **connect** and **fetch the live playbook**.

## How it works (the mental model)

The user ships **only a front-end** — a single self-contained HTML page. There is
no backend to run. Every page carries the **blessed Bailey SDK** (one script tag):
it turns on traffic analytics for free and is the only way anything stateful (a
remembered value, a tally, a form submission) reaches storage. Speak plainly —
"publish it", "give it an address", "it remembers" — never "deploy" or "database".

## Connect

Bailey is an MCP server at `https://<dashboard-host>/mcp` (default
`https://app.trybailey.app/mcp`), JSON-RPC 2.0 over HTTP POST. Auth is the user's
**Supabase Bearer JWT** (the same login as the dashboard) — no key to paste, no
local config. Once connected, the server derives the user's identity and org.

## First action, always: `bootstrap`

Before anything else, call the **`bootstrap`** tool, passing `skill_version: 1`
(the `version` in this file's frontmatter). It returns:

- the **current playbook** — the full sequence, the three primitives, the file rule,
  and the setup reflex (tag the page's type, then relay Bailey's suggestions) — which
  you then **follow**; and
- a freshness check: if it reports `update_required`, tell the user to reinstall the
  latest Bailey skill before continuing.

Everything that can change over time is served (the playbook, the SDK guide via
`get_integration_guide`, the org brand via `get_brand_tokens`, the prompts). Read
those at the moment you need them — never reconstruct their content from memory.

## The one reflex to keep in mind without being told

If a page collects data that can **identify a person** (email, name, phone, a
free-text message), ask the user — *before publishing* — **what it's for** (purpose)
and **how long it's kept** (retention). Those answers are the GDPR register line;
declare them in the `manifest`. A personal data bucket with no purpose is refused.
The playbook details the rest.

## Escape hatch — the dashboard (or the CLI)

Multi-file sites, binary assets, or anything over ~1 MB do not go through MCP.
Send the user to their Bailey **dashboard** to drag-and-drop the folder (the help
center shows how); the `bailey` CLI (`bailey deploy <dir> --app <label>`, `bailey
data`, `bailey erase`) covers advanced setups. The MCP path serves the verb and
the norm, not binary bulk.