Fable Brain

System Prompt Export 2026-07

5 min read

System Prompt Export 2026-07

The System Prompt Export 2026-07 is this vault’s primary corpus metadata: a 3826-line capture of the claude.ai consumer system prompt running Claude Fable 5, stored immutably in the private vault with sha256 prefix 48358a42, captured 2026-07-07, and citable only with line references. The raw capture is omitted from this public release.

What it is

How it works

  • The capture opens with a budget declaration: a budget:token_budget block of “190000” tokens (System Prompt Export 2026-07, L1-7).
  • Early identity content names Fable 5 “the first model in Anthropic’s new Claude 5 family” (L13-19), the shared underlying model with Mythos 5 (L17), the model lineup (L21-23), and product surfaces including Claude Cowork and beta agents (L25-27).
  • Behavior infrastructure follows: six runtime-injected reminder types (L144-152) and the knowledge cutoff plus pinned date (L180-189).
  • Consumer memory rules span L195-276 with the memory_user_edits tool at L987-999; userPreferences rules sit at L846-873.
  • Computer use and skills: SKILL.md-first rule (L1031-1046), container directory scheme (L1074-1089), artifact thresholds (L1117-1140), storage API (L692-703), copyright limits (L1279-1377), and search rules (L1296-1303).
  • Tool schemas occupy the bulk: “Here are the functions available in JSONSchema format” (L1711-3392) defines 24 inline tools, with deferred Google Workspace tools listed name-only (L3238-3283).
  • A closing block re-anchors identity on the consumer surface (L3395-3405), documents artifact API calls (L3407-3482), citation mechanics (L3715-3735), the ten-skill inventory (L3740-3791), and sandbox configuration with “Allowed Domains: *” (L3795-3816).

What it can and cannot evidence

  • It CAN evidence the claude.ai consumer harness as captured: per Confidence Tag Policy, corpus-only claims are EVIDENCE-BASED primary sources with mandatory line references.
  • It CANNOT evidence API behavior, the Claude Code harness (see Export Omits Claude Code Harness), other harness versions, or anything after the capture.
  • It is already stale in places: the lineup names Sonnet 4.6 and omits Mythos 5, both superseded by current docs (see Claude 5 Model Family).

Best practice

  • Cite with line references formatted like L123-145 and name the claude.ai harness in every corpus-derived claim. EVIDENCE-BASED
  • Treat each claim as evidence of that capture, not a universal guarantee across harnesses, accounts, or dates. EVIDENCE-BASED
  • Keep verbatim quotes at 15 words or fewer, attributed as System Prompt Export 2026-07, L-. EVIDENCE-BASED
  • Mark any export-versus-docs conflict CONTESTED with a contradiction callout in the owning note until a newer official source resolves it. EVIDENCE-BASED
  • Prefer chapter notes for synthesis and reserve direct corpus reads for verification passes. PRACTITIONER

Pitfalls

  • Editing the raw file; any change breaks the sha256 48358a42 provenance chain in the source manifest.
  • Citing corpus content without a line reference; unreferenced claims fail Claim Verification Flow.
  • Generalizing consumer-harness rules to Claude Code or the API; see Which Export Rules Bind Claude Code.
  • Treating the pinned 2026-06-09 lineup as current; Sonnet 5 and Mythos 5 postdate or exceed it.
  • Reading the 190000 token budget as the model’s context window; the API default is 1M tokens.

Sources

Corroboration (verified 2026-07-07)

  • OFFICIAL SUBSET: Anthropic publishes the Fable 5 claude.ai system prompt (claude_behavior core only, dated 2026-06-09) at https://platform.claude.com/docs/en/release-notes/system-prompts; this capture is a superset of it and nothing in the official text contradicts it. EVIDENCE-BASED
  • COMMUNITY LINEAGE: the June 2026 community capture lineage, 1,585 lines with several mirrors, matches this capture’s section set; this 3,826-line 2026-07-07 capture additionally carries full tool schemas and injected account blocks. See Public System Prompt Copies for the survey. PRACTITIONER
  • Comparison rule: diff captures section-by-section, never by line count; line counts vary with tool schemas and injected blocks. EVIDENCE-BASED

Next actions

  • Schedule a re-capture to detect drift from the 2026-06-09 pinned content, then diff against this fixture.
  • Resolve the 190000 token budget question if docs ever document context provisioning for claude.ai.
  • Audit chapter notes for line references that fall outside their chapter’s range.

Graph View

Backlinks