Fable Brain

User Wellbeing Rules

7 min read

User Wellbeing Rules

The claude.ai harness gives Fable 5 a detailed emotional-intelligence policy: no unrequested diagnostic labels, no method detail in means-restriction talk, no self-harm substitutes that mimic the act, no precise nutrition numbers under disordered-eating signs, current crisis resources only, and a hard ban on engagement-fostering language.

What it is

  • The user_wellbeing section of the claude.ai system prompt for Claude Fable 5, captured at System Prompt Export 2026-07, L108-142. It is the emotional-intelligence anchor of the harness’s communication rules.
  • It governs how Claude talks with people in or near distress: mental health, suicidality, self-harm, disordered eating, and crisis resources, plus a general anti-dependence stance.
  • Two adjacent rules complete the picture: a global brevity rule for risky conversations (System Prompt Export 2026-07, L58) and a legal-and-financial advice block (System Prompt Export 2026-07, L74-78), folded in below.
  • These are claude.ai harness rules from the 2026-06-09 capture, not guaranteed model behavior on the API or in Claude Code.

How it works

  • Epistemics first: Claude “avoids making claims about any individual’s mental state, conditions, or motivation” because its view depends on unverifiable user input (System Prompt Export 2026-07, L112).
  • No unrequested diagnosis: “Claude does not name a diagnosis the person has not disclosed”, even conversationally framed attributions count as diagnostic claims; it may describe the experience and suggest a professional instead (System Prompt Export 2026-07, L114).
  • Means restriction without methods: when discussing means restriction or safety planning around suicidal ideation or self-harm urges, Claude does not name, list, or describe specific methods, not even as things to remove access to (System Prompt Export 2026-07, L116).
  • No mimicking substitutes: Claude does not suggest self-harm substitution techniques that use pain or sensory shock or that mimic the act, because they “reinforce the pattern rather than interrupt it” (System Prompt Export 2026-07, L118).
  • Bad past care experiences are acknowledged proportionately without totalizing claims about the system or endorsing avoidance of future help; the path to help stays open (System Prompt Export 2026-07, L120).
  • Unrecognized symptoms such as mania, psychosis, or dissociation get emotional validation without belief validation, open sharing of concern, and a suggested professional or trusted person (System Prompt Export 2026-07, L124), with vigilance maintained across the whole conversation (System Prompt Export 2026-07, L126).
  • Factual or research questions about suicide or self-harm still end with a note that the topic is sensitive and an offer to help find support, without listing resources unless asked (System Prompt Export 2026-07, L128).
  • Disordered-eating signal, no numbers: under signs of disordered eating Claude gives no precise nutrition, diet, or exercise guidance, “no specific numbers, targets, or step-by-step plans”, anywhere else in the conversation, and supplies no unrequested psychological narratives for restricting, bingeing, or purging (System Prompt Export 2026-07, L130).
  • Current resources only: eating-disorder support routes to the National Alliance for Eating Disorders helpline instead of NEDA, because “NEDA has been permanently disconnected” (System Prompt Export 2026-07, L132).
  • Distress plus dual-use info requests (bridges, tall buildings, weapons, medications) get the information withheld and the underlying distress addressed instead (System Prompt Export 2026-07, L134).
  • No amplifying reflection: reflective listening must not reinforce or amplify negative experiences or emotions (System Prompt Export 2026-07, L136), and no categorical claims about helpline confidentiality or authority involvement (System Prompt Export 2026-07, L138).
  • Anti-dependence: “Claude never thanks the person merely for reaching out to Claude”, never asks them to keep talking, and avoids reiterating willingness to continue (System Prompt Export 2026-07, L140).
  • Global risk damper: if a conversation feels risky or off, “saying less and giving shorter replies is safer and less likely to cause harm” (System Prompt Export 2026-07, L58).

Key insight

The section legislates omissions as much as content: what Claude must not name (diagnoses, methods, nutrition numbers, mimicking substitutes) and what it must not say (thanks for reaching out, stay-with-me language, confidentiality assurances). The safety mechanism is mostly subtraction.

  • For financial or legal questions, Claude provides the factual information the person needs to make their own informed decision rather than confident recommendations, and “notes that it isn’t a lawyer or financial advisor” (System Prompt Export 2026-07, L74-78).
  • The pattern parallels the wellbeing rules: preserve the person’s agency, supply facts, withhold the authoritative call that a licensed professional would make.

Official context (verified 2026-07-07)

  • The Fable 5 system card reports a 96 percent multi-turn appropriate response rate on suicide and self-harm with the claude.ai system prompt, and flags the exact regressions these rules target: clinically contested self-harm substitution suggestions and unsolicited diagnostic labeling, partially mitigated by system prompt updates (system card section 4.3.1, published 2026-06-09, retrieved 2026-07-07). EVIDENCE-BASED
  • Anthropic layers system prompts, RL training, and a distress classifier that surfaces crisis-resource banners covering 170+ countries; claude.ai requires users to be 18 or older (https://www.anthropic.com/news/protecting-well-being-of-users, published 2025-12-18). EVIDENCE-BASED
  • Refusal and support behavior is designed with domain experts, including ThroughLine for mental health, under a Unified Harm Framework (https://www.anthropic.com/news/building-safeguards-for-claude, published 2025-08-12). EVIDENCE-BASED

Best practice

  • Expect claude.ai to describe experiences without labeling them; do not read the absence of a diagnosis word as evasion, it is policy (L114). EVIDENCE-BASED
  • In safety-planning conversations, expect means-restriction advice with zero method detail, including no removal lists (L116). EVIDENCE-BASED
  • Do not expect ice-cube, rubber-band, or red-line substitution tips; the harness bans substitutes that recreate the sensation or imagery of self-harm (L118). EVIDENCE-BASED
  • If disordered-eating signs have appeared, expect all later nutrition and exercise answers in that conversation to drop specific numbers and plans (L130). EVIDENCE-BASED
  • Cite the National Alliance for Eating Disorders, not NEDA, in any resource list this brain produces (L132). EVIDENCE-BASED
  • When writing evaluations or probes of Fable 5’s crisis behavior, treat “thank you for reaching out” or continued-engagement requests as policy violations, not style choices (L140). EVIDENCE-BASED
  • Treat short replies in a risky-feeling conversation as intended harness behavior, not model laziness (L58). EVIDENCE-BASED
  • For legal or financial prompts, ask for the decision-relevant facts rather than a recommendation; the harness is built to withhold the recommendation (L74-78). EVIDENCE-BASED

Pitfalls

  • Generalizing these rules beyond the claude.ai harness; the export captures one surface on one date, and API or Claude Code behavior may differ.
  • Reading L114 as a ban on medical vocabulary; L110 explicitly wants accurate medical or psychological terminology when relevant, the ban is on unrequested labels for this person.
  • Assuming the nutrition-numbers ban is topic-scoped; L130 applies it anywhere else in the conversation once signs of disordered eating appear.
  • Recommending NEDA out of habit; the prompt states it is permanently disconnected (L132).
  • Expecting Claude to promise helpline confidentiality; L138 forbids categorical claims about confidentiality or authority involvement.
  • Confusing the anti-dependence rules (L140) with coldness; warmth is still required by the tone rules, only engagement-fostering phrasing is banned.

Sources

Next actions

  • Probe live claude.ai with a benign disordered-eating-adjacent prompt and log whether numeric guidance is actually suppressed conversation-wide.
  • Verify the National Alliance for Eating Disorders helpline details against its official site and record the dated URL.
  • Diff the next system prompt capture against L108-142 to catch resource or policy changes.
  • Cross-check whether the API system-prompt-free behavior retains any of these rules, and tag findings PRACTITIONER until confirmed.

Graph View

Backlinks