Fable Brain

Which Export Rules Bind Claude Code

4 min read

Which Export Rules Bind Claude Code

Unresolved: the export is a claude.ai capture, so none of its rules are proven to bind Claude Code, whose behavior is documented separately at code.claude.com.

What it is

  • An open question left behind by Corpus Scope Decision: the corpus contains rich behavioral rules, but every one of them is attested only for the claude.ai harness.
  • Candidate rules whose reach is unknown include the copyright hard limits (quotes over 15 words as severe violations, one quote per source, absolute lyrics ban, System Prompt Export 2026-07, L1279-1377), citation mechanics (L3715-3735), search effort scaling and the unrecognized-entity search mandate (L180-189, L1296-1303), artifact usage thresholds (L1117-1140), and the six runtime-injected reminder types with the rule that reminders never reduce restrictions (L144-152).
  • The question matters because operators reading the export may wrongly assume Claude Code enforces the same limits, or wrongly assume it enforces none of them.

How it works

What we know

  • The corpus is a June 9, 2026 claude.ai consumer prompt snapshot; nothing in it describes the Claude Code harness (see Export Omits Claude Code Harness).
  • Claude Code has its own documented rule surface: CLAUDE.md memory, permission modes, settings, hooks, subagents, and skills, canonical at code.claude.com after the 301/308 redirects (https://code.claude.com/docs/en/overview).
  • The public Claude Code docs verified on 2026-07-07 do not republish the export’s copyright, citation, search, or artifact rules, so no documented overlap is established.
  • One concrete tension exists: the corpus shows claude.ai container egress with a wildcard domain allowlist (L1074-1089, L3795-3816), while public Claude Code sandbox guidance warns that broad allowlists enable exfiltration; the docs describe a different surface, so this is a divergence signal, not a contradiction.

What remains unknown

  • Whether shared policy layers (for example copyright limits) are injected into Claude Code sessions by a prompt we have never captured.
  • Whether model-level training makes some export rules behave identically everywhere even without prompt text.

What would resolve it

  • A captured Claude Code system prompt of comparable fidelity to the claude.ai export.
  • An official Anthropic statement mapping which policy layers are shared across harnesses.
  • Behavioral probes: identical quote-length, citation, and artifact-threshold tasks run in both harnesses and diffed.

Best practice

  • Treat every export rule as claude.ai-only until a probe or capture proves otherwise. PRACTITIONER
  • Source all Claude Code behavior claims from code.claude.com pages, which the redirect chain establishes as canonical (https://code.claude.com/docs/en/overview). EVIDENCE-BASED
  • When writing cross-harness guidance, mark the harness on every claim, per Corpus Scope Decision. PRACTITIONER
  • Design probes with per-rule falsifiable predictions instead of one vague “same rules” test. PRACTITIONER

Pitfalls

  • Assuming a rule as fundamental-sounding as the copyright quote limit must be universal; it is attested only at L1279-1377 of a claude.ai capture.
  • Assuming the inverse, that Claude Code is unconstrained, because its docs are silent on content rules.
  • Treating the sandbox allowlist tension as a contradiction; the two sources describe different surfaces.
  • Letting this question rot: the corpus snapshot ages while Claude Code ships weekly, so the divergence only grows.

Sources

Next actions

  • Design the cross-harness quote-length probe (pairs with Tool Search Behavior Probe methodology).
  • Watch code.claude.com docs for any content-policy page that would document overlap.
  • Upgrade answer_quality past draft once one probe has run in both harnesses.

Graph View

Backlinks