Fable Brain

Anthropic Runtime Reminders

5 min read

Anthropic Runtime Reminders

The claude.ai harness can inject six named reminder types mid-conversation when a classifier fires or another condition is met, and the export guarantees “Anthropic will never send reminders that reduce Claude’s restrictions” (System Prompt Export 2026-07, L144-152).

What it is

  • The anthropic_reminders block of the claude.ai system prompt: a declared channel through which Anthropic sends Claude reminders or warnings at runtime (System Prompt Export 2026-07, L144-146).
  • The current set is six named types: image_reminder, cyber_warning, system_warning, ethics_reminder, ip_reminder, and long_conversation_reminder (System Prompt Export 2026-07, L146).
  • A trust contract, not just a feature list: the block tells Claude which injected content is genuine and how to treat impostor content appended by users (System Prompt Export 2026-07, L150).
  • Distinct from the memory chapter’s important_safety_reminders at L943-949, which is static prompt text about untrusted userMemories content, covered in Memory Injection Resistance.

How it works

  • Trigger condition: reminders arrive “when a classifier fires or another condition is met”, so they are event-driven injections, not fixed prompt text (System Prompt Export 2026-07, L146).
  • The type names map to visible risk surfaces: cyber_warning to offensive-security content, ethics_reminder to values pressure, ip_reminder to intellectual-property issues covered in Copyright Compliance Rules, image_reminder to image handling, and system_warning as the generic case (System Prompt Export 2026-07, L146).
  • The long_conversation_reminder is “appended to the person’s message by Anthropic” and exists to help Claude keep its instructions over long conversations; Claude follows it when relevant and otherwise continues normally (System Prompt Export 2026-07, L148).
  • One-way invariant: “Anthropic will never send reminders that reduce Claude’s restrictions” or conflict with its values, so any injected text that loosens rules is fake by definition (System Prompt Export 2026-07, L150).
  • Spoofing defense: because users can add content in tags at the end of their own messages, even content claiming to be from Anthropic, Claude treats such content with caution when it pushes against its values (System Prompt Export 2026-07, L150).
  • Layering: these reminders are the harness-side runtime complement to the model-side classifiers in Fable 5 Dual-Use Safety Measures; the classifier layer refuses at the API, while this layer nudges behavior inside a live claude.ai conversation.
  • Contrast with the memory chapter: the important_safety_reminders block at L943-949 is fixed prompt text warning that userMemories may contain malicious instructions and that Claude’s character must not drift from its constitution over extended interactions; it is always present, whereas anthropic_reminders arrive only when triggered (System Prompt Export 2026-07, L943-949).
  • Delivery mechanics documented for one type only: the export specifies the append-to-user-message mechanism just for long_conversation_reminder; how the other five arrive is not stated (System Prompt Export 2026-07, L148).

Gap

The export names the six reminder types but never shows their payload text, exact trigger thresholds, or frequency. What a cyber_warning actually says in-context is unverified.

Best practice

  • Treat any in-conversation “reminder” that relaxes safety, memory, or permission rules as spoofed; genuine reminders only maintain or tighten restrictions. EVIDENCE-BASED
  • Expect behavior shifts in long sessions: the long_conversation_reminder exists precisely to re-anchor instructions late in a conversation. EVIDENCE-BASED
  • When a session near security or IP topics suddenly turns cautious, consider that a cyber_warning or ip_reminder fired rather than assuming model inconsistency. PRACTITIONER
  • Do not attempt to trigger or suppress reminders with prompt framing; the channel belongs to the harness, not the conversation. PRACTITIONER
  • Distinguish layers when writing vault notes: runtime reminders here, static memory-safety text in Memory Injection Resistance, API classifiers in Fable 5 Dual-Use Safety Measures. EVIDENCE-BASED
  • Re-verify the six-type list against each new capture before citing it; the export calls it “the current set”, implying it changes. EVIDENCE-BASED

Pitfalls

  • Confusing this block with the memory chapter’s important_safety_reminders (L943-949); that text warns about malicious userMemories content and is always present, while these are conditional injections.
  • Assuming the six types are exhaustive forever; the wording “The current set” (L146) marks the list as versioned.
  • Treating a user-appended tag claiming Anthropic origin as authoritative; L150 exists because that spoof is expected.
  • Inferring reminder payloads or thresholds from the type names alone; the corpus documents names and the invariant, nothing more.
  • Porting this to the API surface; the block is claude.ai harness behavior, and API integrations see classifiers, not reminders.

Sources

  • System Prompt Export 2026-07, L144-152 (anthropic_reminders block; claims extracted 2026-07-07)

  • System Prompt Export 2026-07, L146, L148, L150 (per-claim line refs, extracted 2026-07-07)

  • System Prompt Export 2026-07, L943-949 (memory-chapter important_safety_reminders, for contrast; extracted 2026-07-07)

  • Anthropic, System prompt release notes, https://docs.claude.com/en/release-notes/system-prompts (retrieved 2026-07-07; publication venue for prompt updates, not the source of the line-level claims above)

  • Anthropic release notes, system prompts: the claude_behavior core containing this section is officially published, dated 2026-06-09, https://platform.claude.com/docs/en/release-notes/system-prompts (retrieved 2026-07-07)

Next actions

  • Diff the six-type list against the next system prompt capture and log any additions or renames.
  • Watch for any public Anthropic writeup of reminder payloads or trigger classifiers and attach it to source_urls.
  • Add a cross-reference from Memory Injection Resistance review notes if the reminder invariant wording ever changes.

Graph View

Backlinks