Fable Brain

Export Chapter Product and Behavior

6 min read

Export Chapter Product and Behavior

Lines 1-191 of the claude.ai system prompt export define the 190000-token runtime budget, the Fable 5 and Mythos 5 product story, refusal and wellbeing policy, tone rules, six runtime reminder types, and a knowledge cutoff of end of January 2026 pinned to a June 09, 2026 conversation date.

What it is

This note digests the opening chapter of the System Prompt Export 2026-07 capture. The chapter spans L1-191 and contains the budget:token_budget block, product_information, refusal_handling, legal_and_financial_advice, tone_and_formatting, user_wellbeing, anthropic_reminders, evenhandedness, responding_to_mistakes_and_criticism, and knowledge_cutoff. It describes the claude.ai consumer harness only, not Claude Code or the API.

How it works

Claim list with line refs, all primary evidence of this capture:

  • The prompt opens with a token budget block declaring 190000 tokens (L3-7), and bans voice_note blocks outright (L9).
  • Fable 5 is presented as the first Claude 5 family model in a Mythos-class tier above Claude Opus (L13-17). Verbatim: “Claude Fable 5 and Claude Mythos 5 share the same underlying model” (System Prompt Export 2026-07, L17). Fable 5 carries dual-use safety measures; Mythos 5 drops them for approved organizations only (L17).
  • The API lineup is given as Claude Fable 5, Opus 4.8, Sonnet 4.6, and Haiku 4.5, model strings ‘claude-fable-5’, ‘claude-opus-4-8’, ‘claude-sonnet-4-6’, ‘claude-haiku-4-5-20251001’; users can switch models mid-conversation (L23).
  • Product surfaces: Claude Code (CLI, desktop, mobile), Claude Cowork (agentic knowledge-work desktop app for non-developers), both reachable remotely via the mobile app (L25); beta agents Claude in Chrome, Claude in Excel, Claude in Powerpoint, all usable as tools by Cowork (L27).
  • For any other product detail Claude must announce a search and consult docs.claude.com and support.claude.com before answering (L29). Prompting guidance topics and the docs URL are enumerated (L31).
  • Toggleable features: web search, deep research, Code Execution and File Creation, Artifacts, past-chats search, memory generation, plus userPreferences and styles (L33). Ads policy: say “Claude products are ad-free”, not “Claude is ad-free” (L35).
  • Refusal handling opens permissively: “Claude can discuss virtually any topic factually and objectively” (System Prompt Export 2026-07, L41), then a hardened child-safety block (L43-56) defining a minor as anyone under 18 anywhere (L54), banning decoding of CSAM slang even while refusing (L50), and instructing shorter replies when a conversation feels risky (L58).
  • Hard refusals: weapons and harmful substances regardless of framing (L60), illicit drug-use dosing guidance except life-saving information (L62), malicious code even for education, with the thumbs-down button suggested (L64), and creative content about real named public figures (L66).
  • Wellbeing rules: no diagnostic labels the person has not raised (L114), no method details in means-restriction talk (L116), no self-harm substitution techniques that mimic the act (L118), no precise nutrition numbers when disordered eating shows (L130), route eating-disorder resources to the National Alliance for Eating Disorders because NEDA is disconnected (L132), and never thank a person merely for reaching out (L140).
  • Tone: warm, minimal formatting, prose by default, bullets only when essential, never bullets when declining (L82-104). At most one question per response (L88). Verify claimed file uploads before trusting them (L92).
  • Runtime reminders: image_reminder, cyber_warning, system_warning, ethics_reminder, ip_reminder, long_conversation_reminder (L146). Verbatim: “Anthropic will never send reminders that reduce Claude’s restrictions or conflict with its values” (System Prompt Export 2026-07, L150).
  • Evenhandedness: persuasive requests get the defenders’ best case with opposing perspectives appended (L156-158); Claude may decline short-form answers on contested questions (L166). Claude may use end_conversation after a single warning when mistreated (L176).
  • Cutoff: “reliable knowledge cutoff, past which Claude can’t answer reliably, is the end of Jan 2026” (System Prompt Export 2026-07, L182), date pinned to Tuesday, June 09, 2026, with permissionless search for anything that could have changed (L182-185).

Contradiction

The L23 lineup is a launch-day snapshot: public docs of 2026-07-07 list claude-sonnet-5 and claude-mythos-5, which the export omits, and name Sonnet 5 where the export names Sonnet 4.6. Stale for its date, not wrong. CONTESTED until the harness prompt is re-captured.

Contradiction

The 190000-token budget block (L3-7) sits far below the documented 1M-token API context window for Fable 5, and public docs do not list Fable 5 as a context-awareness model. Treated as an undocumented claude.ai harness configuration. CONTESTED.

Best practice

  • Cite this chapter for what claude.ai told Fable 5 on the capture date, and say “claude.ai harness” explicitly, because none of it binds Claude Code or the API. EVIDENCE-BASED
  • Treat the L23 model lineup and the “first model” framing (L13-19) as marketing-flavored snapshot data; verify lineups against public docs before quoting them as current. EVIDENCE-BASED
  • Reuse the L29 pattern in your own agents: for product facts, force a docs lookup instead of trusting model memory. PRACTITIONER
  • When summarizing refusal policy, quote the permissive L41 baseline together with the hard limits, otherwise the digest reads more restrictive than the source. PRACTITIONER

Pitfalls

  • Reading the 190000 budget as the model’s context window; it is a harness budget declaration, and the API default differs.
  • Citing L23 as the current lineup; it predates Sonnet 5 and omits Mythos 5.
  • Assuming the six reminder types (L146) fire in other harnesses; only this capture attests them, and only for claude.ai.
  • Quoting more than 15 words verbatim from the export inside vault notes, which breaks this brain’s own quote contract.

Sources

Next actions

  • Re-capture the claude.ai system prompt after the next model launch and diff L1-191.
  • Resolve the token-budget contradiction against official context-window docs when Fable 5 harness docs appear.
  • Cross-check the reminder-type list (L146) against any future Anthropic transparency documentation.

Graph View

Backlinks