Fable Brain

Export Chapter Artifacts API and Citations

5 min read

Export Chapter Artifacts API and Citations

Lines 3407-3826 of the export let artifacts call the Anthropic API with harness-handled auth pinned to a Sonnet 4 snapshot, define cite-tag mechanics, inventory ten skills, and reveal a wildcard network allowlist, five read-only mounts, and thinking_mode auto.

What it is

This note digests the final chapter of the System Prompt Export 2026-07, spanning L3407-3826: anthropic_api_in_artifacts, citation_instructions, the user location line, the skills inventory, network_configuration, filesystem_configuration, the thinking_mode tag, and the start of the Human turn with a userPreferences placeholder.

How it works

Claim list with line refs:

  • Artifacts can call the standard /v1/messages endpoint directly, a capability users may call “Claude in Claude”, “Claudeception”, or AI-powered artifacts (L3411). Auth is invisible: “The assistant should never pass in an API key, as this is handled already” (System Prompt Export 2026-07, L3417).
  • The example call pins model to claude-sonnet-4-20250514 with the comment to always use Sonnet 4, and fixes max_tokens at 1000 (L3426-3427), regardless of Fable 5 running the outer conversation.
  • Responses return mixed content blocks (text, tool_use, tool_result, image, document) that must be filtered by type, never by position (L3437-3448, L3492-3513). Structured output is prompted as JSON-only with fences stripped before parsing (L3453-3457, L3689-3696).
  • Artifact API calls accept an mcp_servers parameter, illustrated with Asana, Gmail, and Salesforce (L3464), and the prompt embeds the user’s live connector list: Google Drive, Gmail, Google Calendar, Canva, and Figma (L3482). Web search is enabled inside artifacts via the web_search_20250305 tool type (L3548-3560), and MCP plus search can be combined (L3566).
  • The artifact-side model is stateless between completions, so full history and complete application state must ride in every request (L3636-3681). React artifacts must never use HTML form tags, only event handlers (L3705-3709).
  • Citations: every specific claim following from web_search results gets cite tags whose index encodes document and sentence spans, DOC_INDEX-START:END, comma-separated for multiple sections (L3719-3723), using the minimum sentences needed (L3725), and never citing document_context (L3727). Verbatim: “Claims must be in your own words, never exact quoted text” (System Prompt Export 2026-07, L3729).
  • The user’s approximate location (Reykjavík, Capital Region, IS in this capture) is injected in plain text (L3738).
  • The skills inventory lists ten skills with filesystem locations: docx, pdf, pptx, xlsx, product-self-knowledge, frontend-design, file-reading, and pdf-reading under /mnt/skills/public, plus learn and skill-creator under /mnt/skills/examples (L3740-3791).
  • product-self-knowledge overrides training memory for Anthropic product facts: “Any time you would otherwise rely on memory for Anthropic product details, verify here instead” (System Prompt Export 2026-07, L3757).
  • Network config for bash_tool: Enabled true, Allowed Domains *, with an x-deny-reason header explaining egress failures (L3797-3801). Filesystem config mounts five directories read-only: /mnt/user-data/uploads, /mnt/transcripts, and the three /mnt/skills trees (L3807-3812).
  • thinking_mode is set to auto (L3818), and the Human turn opens with a userPreferences placeholder (L3820-3826).

Contradiction

The pinned claude-sonnet-4-20250514 (L3426) is absent from the documented 2026-07-07 lineup (claude-fable-5, claude-mythos-5, claude-opus-4-8, claude-sonnet-5, claude-haiku-4-5). The pin may still be served but is stale relative to public docs, plausibly a deliberate cost and latency choice. CONTESTED.

Contradiction

The wildcard allowlist (L3799) sits in tension with public Claude Code sandbox guidance warning that broad allowlists enable exfiltration; the docs describe a different surface, so this is tension, not direct conflict. CONTESTED.

Best practice

  • Build AI-powered artifacts against /v1/messages with no key handling and type-filtered block parsing, exactly as specified (L3417, L3492-3513). EVIDENCE-BASED
  • Pass full state in every artifact-internal API call; the inner model has no memory between completions (L3638). EVIDENCE-BASED
  • Follow the paraphrase-only citation rule in any cited answer: cite tags grant attribution, not reproduction (L3729). EVIDENCE-BASED
  • Route Anthropic product questions through product-self-knowledge style verification instead of model memory (L3756-3758). EVIDENCE-BASED
  • Do not assume the Sonnet 4 pin performs like Fable 5; scope inner-model expectations to a small, cheap model. PRACTITIONER

Pitfalls

  • Passing an API key inside artifact code; the harness already injects auth (L3417).
  • Raising max_tokens above 1000 expecting effect; the export fixes it (L3427).
  • Reading data.content[0] positionally; block order is not guaranteed (L3492-3496).
  • Citing sentences from document_context, which the rules exclude (L3727).
  • Generalizing the wildcard egress allowlist to Claude Code, where sandbox policy differs.

Sources

Next actions

  • Test whether the Sonnet 4 pin still resolves from a live artifact and log the response model.
  • Re-capture after the next artifacts update; the mcp_servers connector list is per-user and volatile.
  • Verify the wildcard egress claim by fetching a novel domain from bash_tool in a live session.

Graph View

Backlinks