Fable Brain

Export Omits Claude Code Harness

5 min read

Export Omits Claude Code Harness

The System Prompt Export 2026-07 contains no Claude Code system prompt, tool schemas, or harness rules, so every Claude Code claim in this vault rests on web documentation, never on the corpus.

Gap

The corpus covers claude.ai only. There is no captured Claude Code system prompt of comparable fidelity, which means the vault’s primary-source evidence class (rule 1 of the confidence policy) is unavailable for the entire claude-code-harness domain. All Claude Code claims are corroborated instead through code.claude.com documentation and dated practitioner posts.

What it is

  • A gap register entry documenting the corpus’s coverage boundary.
  • Positive evidence of the boundary: the corpus inventory is consumer-shaped end to end, including a 24-tool inline harness inventory of consumer widgets (System Prompt Export 2026-07, L1711-3392), the claude.ai container layout with /mnt/user-data paths (L1074-1089), consumer memory mechanics (L195-276), the claude.ai skills surface under /mnt/skills (L3740-3791), and an identity block restating the June 09, 2026 date (L3395-3405). None of it describes the Claude Code harness.
  • Further consumer-only inventory reinforces the boundary: deferred-tool plumbing behind tool_search (L2940-3283), userPreferences application rules (L846-873), and the artifact usage thresholds (L1117-1140) all belong to the claude.ai surface.
  • The complementary decision, Corpus Scope Decision, turns this observation into a binding tagging rule.

How it works

  • Consequence for evidence classes: claude.ai notes can cite primary capture lines; Claude Code notes cite only rule 2 sources, chiefly code.claude.com pages, whose canonical status is established by the 301 redirects from docs.claude.com and the 308 redirect from the former engineering best-practices post (https://code.claude.com/docs/en/overview).
  • Consequence for open questions: whether export rules (copyright, citation, search, artifacts) bind Claude Code is undecidable from current evidence; tracked in Which Export Rules Bind Claude Code.
  • Consequence for freshness: Claude Code documentation reflects a fast release train (auto memory default since v2.1.59, auto mode v2.1.83+, /agents wizard removed v2.1.198), while the corpus is frozen at June 9, 2026, so the two evidence pools age at different speeds.
  • Consequence for tooling notes: tool_search and deferred-tool claims are claude.ai evidence, so Claude Code tool behavior must be sourced from its own documentation pages.
  • Consequence for automation notes: the documented non-interactive surface (claude -p, —bare) exists only in web sources, so Headless Claude Code and CI carries no primary capture either.
  • The gap does not lower confidence on documented Claude Code claims; rule 2 evidence remains EVIDENCE-BASED. Only the primary capture class is missing.
  • The gap closes only with a Claude Code prompt capture or an official publication of its system prompt and tool schemas.

Best practice

  • Never cite corpus line references for Claude Code behavior; use code.claude.com pages with retrieval dates. PRACTITIONER
  • Record Claude Code version numbers (v2.1.59, v2.1.83, v2.1.198) alongside claims, since docs describe a moving target. PRACTITIONER
  • Treat practitioner posts on Claude Code as rule 3 sources unless they cite official docs, then verify the doc directly. PRACTITIONER
  • Revisit this gap at every brain refresh; a public prompt capture would reshape the whole claude-code-harness domain. PRACTITIONER
  • Mirror the corpus-side chapter structure when a Claude Code capture arrives so the two harnesses stay comparable note-for-note. PRACTITIONER

Pitfalls

  • Silently borrowing claude.ai container facts (uploads, outputs, /mnt paths) into Claude Code notes; the surfaces differ.
  • Assuming the corpus omission means Claude Code has no system prompt; it means we have not captured one.
  • Citing the GitHub repository as behavioral documentation; it anchors the project but the docs pages carry the behavior claims.
  • Forgetting that the wildcard egress allowlist detail (L3795-3816) is claude.ai evidence, despite its relevance to Claude Code sandbox debates.
  • Citing the claude.ai skills surface under /mnt/skills as evidence about Claude Code skills; the two skill systems are documented separately.
  • Treating this gap as permanent; harness prompts change, and a future export may include Claude Code, so this register entry carries a standing watch duty.

Sources

Partial fill available (found 2026-07-07)

  • The Claude Code prompt fragments are publicly extracted per release at link omitted for public release (v2.1.202, 2026-07-06, 515+ prompt strings including a Fable 5 model identity fragment). Ingesting a pinned release into the immutable source layer via the Corpus Ingestion Flow would close most of this gap; until then the gap stands. PRACTITIONER

Next actions

  • Watch for any credible Claude Code system prompt capture and ingest it as a new corpus.
  • Add version-number fields to claude-code-harness notes at the next lint pass.
  • Re-check the docs redirect chain at the next brain refresh to confirm canonical status.

Graph View

Backlinks