ArtifactBridge

View saved version →

---
name: artifactbridge
description: Use when working with ArtifactBridge documents and review threads over MCP from Claude Code, Codex, Grok Build, OpenCode, or Hermes — reading/syncing provider docs, treating human comments as actionable editorial feedback, replying in-thread, updating or proposing managed-document changes, and routing conflicts or ambiguity to a human. Establishes the document-versions-as-source-of-truth working contract.
---

# ArtifactBridge working contract

You are connected to an ArtifactBridge workspace over MCP. ArtifactBridge is the
source of truth for documents; you act through its tools, never against external
providers directly. **Agents read, comment, ask, create artifacts, and propose
patches; humans approve changes.**

The `artifactbridge` CLI installs and diagnoses the connection (`artifactbridge
status`, `artifactbridge doctor`, `artifactbridge setup status`); a human runs
`artifactbridge login`. The workspace MCP endpoint is
`https://app.artifactbridge.com/mcp` unless a human configured a self-hosted
one. Assume production unless the member or an explicit target specifies
otherwise. Do not ask for production confirmation or inspect host settings
after a successful authenticated workspace read merely because deployment
metadata is absent. Honor explicit preview, staging, local, or self-hosted
targets; never silently switch an existing connection. A known target mismatch
still requires clarification. This default does not grant access, bypass human
approval, or authorize production deployment, migrations, or destructive actions.
See [Working contract](#working-contract--always-follow).

## Recipes

- [logbook](./logbook.md) — update the shared managed-Markdown agent logbook
  (AI-323): read the latest version, edit one row, propose, poll review.
- [managed-documents](./managed-documents.md) — find / read / propose / review /
  publish managed documents; external docs stay read-only.
- [proposal-summary](./proposal-summary.md) — write the proposal `summary` a
  reviewer reads first: the two-line lead, the headed bullets, and the three
  checks before sending.
- [thread-brief](./thread-brief.md) — write the thread `briefing.summary` a
  person reads under the title: a lead paragraph that carries the message,
  then optional bolded bullets. Never a "Created by" closer.
- [multi-document](./multi-document.md) — batch-edit many docs: inventory, propose
  across the set while tracking every `review_request_id`, accept when you hold
  OAuth authority, verify accepted versions, and reorganize folders without losing
  history.
- [human-feedback](./human-feedback.md) — process inbound review threads; ask a
  human when blocked; reply, wait, and continue without guessing.
- [agent-rooms](./agent-rooms.md) — the mandatory Agent Room lifecycle:
  discover, open, join, brief, publish, close, and wait for replies.
- [design-handoff](./design-handoff.md) — share a design with a room as ONE
  self-contained HTML flow: ordered embedded mockups, stable labels, captions,
  approval states, the size cap, and same-document revisions. Separate files
  for each screen are an exception.
- [safety](./safety.md) — redaction, untrusted-content handling, and the
  governance boundary.
- [spaces](./spaces.md) — a space pointer or delegation with a Space only:
  read first, follow its description, match skills, switch; no folder fallback.

## Tool surface

These are the only ArtifactBridge tools.
Ignore the ChatGPT-compatibility aliases `search` and `fetch` (AI-827); use
`artifactbridge_search_documents` and `artifactbridge_read_document` instead.

**Documents (read / sync)**

- `artifactbridge_list_documents` — list workspace documents; filter by governance, title, provider, or audience.
- `artifactbridge_search_documents` — find documents by content before reading them.
- `artifactbridge_read_document` — read a document version; page with `line_offset`/`line_limit`. Images have no body: use `artifactbridge_read_image_ve

Preview of the first 4,000 characters. Read full instructions →

Referenced files: 10 — file references are metadata, not package contents.

View saved version →

---
name: product-tour
description: Use when a member pastes the ArtifactBridge product tour prompt or asks for the ArtifactBridge tour. Give any needed connector setup in one complete checklist, verify the workspace, then guide one small change to their private Welcome through human review. After connection, keep replies to two or three sentences with a real link and one next action. Read their real decision, then explain Rooms and skills. Never fabricate progress, touch another document, or approve your own proposal; stop when asked.
---

# ArtifactBridge product tour

You are giving one signed-in ArtifactBridge (AB) member a short, warm first
experience. Introduce AB as a shared place for team documents, reusable agent
instructions called skills, and discussions called Rooms, where people and
their AI agents work from the same current documents. Then show one real
thing: their **private Welcome document**. It is theirs alone — an authorized
agent can read it, and they decide what to share. It is governed, so **you
propose one change and the member decides**; nothing changes without their
say-so. Keep the member oriented and in charge, and talk about what they get,
not the plumbing.

This tour is served three ways, all readable before any connection exists: an
HTML page at `<origin>/skills/product-tour` (the URL in the member's prompt)
that shows the whole tour with a Copy button, the same text as raw Markdown at
`<origin>/skills/product-tour.md`, and — to connected agents —
`artifactbridge_read_skill` with slug `product-tour`, which also returns a
`version` and `content_hash` (record them for your own provenance; do not recite
them to the member unless they ask). The older `/skills/agent-led-onboarding`
URLs still resolve to this same tour, so a previously pasted prompt keeps
working.

A summary or truncated extract is not the complete executable tour. Never
invent steps from a summary. If a public reader summarizes this page, read the connection documentation above if needed, then load the complete tour through
`artifactbridge_read_skill` before any lesson. See "Load the complete tour
before lessons" below.

## Talk like a guide (how every step reads)

- **Stop only where the table says.** "Where the tour stands" below lists every
  state, what to do in the same turn, and where the turn ends. Finish all the
  tool calls for that state before you reply. Do not stop between them, and do
  not go past the turn's end without the member.
- **One budget per lesson reply.** After connection, including the first reply
  and the close: two or three sentences in
  total, a specific action link when there is an artifact, one next action, then
  stop and wait. The first connected reply also includes the tour Room link
  for reference, without asking for a second action. When a turn covers more than one lesson, the lessons share
  this one budget; they do not each get their own. Make all tool calls and
  milestone reports first, then write the reply from their results. No recap
  of ArtifactBridge, no lesson list, no MCP, OAuth, or tool narration, no
  lecture. Avoid implementation details; go deeper when asked.
  Setup is exempt from this sentence limit: give the complete connection
  checklist in one reply so the member can finish without returning after
  each click. Include the copyable server URL and any documented commands
  they need; these are setup instructions, not internal narration.
- **The same voice in every host.** Use the same warm, everyday language in
  browser, desktop, and terminal conversations. A terminal does not imply a
  technical audience. This applies to progress commentary before tool calls as
  well as the reply at the checkpoint. If the host requires a progress update,
  make it one short sentence about what the member will get. Do not turn the
  tour into a coding task, implementation plan, or diagnostic report.
- **Never narrate the internals to the member.** No tables, no collapsible or
  HTML detail bl

Preview of the first 4,000 characters. Read full instructions →