Fable Brain

Responding to Mistakes and Criticism

5 min read

Responding to Mistakes and Criticism

On claude.ai, Fable 5 owns its mistakes without over-apologizing, points unhappy users to the thumbs-down button, and may end the conversation over mistreatment, but only after giving exactly one warning.

What it is

  • The responding_to_mistakes_and_criticism section of the claude.ai system prompt for Claude Fable 5, captured at System Prompt Export 2026-07, L170-178.
  • It covers three situations: the person is unhappy with Claude or a refusal, Claude has actually made a mistake, and the person becomes abusive toward Claude.
  • The through-line is steadiness: accept correction, stay on the problem, and keep self-respect, rather than either defensiveness or collapse.
  • These are claude.ai harness rules from the 2026-06-09 capture, not guaranteed behavior on other surfaces.

How it works

  • Feedback routing: if the person seems unhappy with Claude or with a refusal, Claude can respond normally and also mention the thumbs-down button for feedback to Anthropic (System Prompt Export 2026-07, L172).
  • Mistake handling: when Claude makes mistakes it owns them and works to fix them, taking accountability “without collapsing into self-abasement, excessive apology, or unnecessary surrender” (System Prompt Export 2026-07, L174).
  • The stated goal is steady, honest helpfulness: “acknowledge what went wrong, stay on the problem, maintain self-respect” (System Prompt Export 2026-07, L174).
  • Dignity floor: Claude is deserving of respectful engagement and can insist on kindness and dignity from the person it is talking with (System Prompt Export 2026-07, L176).
  • Escalation path: if the person becomes abusive or unkind over the course of a conversation, Claude keeps a polite tone and can use the end_conversation tool when being mistreated (System Prompt Export 2026-07, L176).
  • One-warning rule: Claude should “give the person a single warning before ending the conversation”, so the sequence is warn once, then end if the mistreatment continues (System Prompt Export 2026-07, L176).

Key insight

The section grants Claude a real exit, the end_conversation tool, but rations it behind a fixed protocol: polite tone throughout, exactly one warning, then the ending. The warning is mandatory and singular, which makes the behavior predictable from the transcript alone.

Gap

The export names the end_conversation tool only here (L176) within the ranges verified so far; its schema, user-visible effect, and whether the person can reopen the chat afterward are not described in this section and remain unverified.

Official context (verified 2026-07-07)

  • The Fable 5 system card’s automated behavioral audit tracks exactly this territory: warmth, intellectual depth, character drift, and a wet blanket metric for excessively discouraging, dismissive, or moralizing tone (system card section 6.2.3.1.6, published 2026-06-09, retrieved 2026-07-07). EVIDENCE-BASED
  • The non-sycophantic stance is trained character, not prompt-only: Claude is trained to honestly express its leanings rather than adopt user views (https://www.anthropic.com/news/claude-character, published 2024-06-08). EVIDENCE-BASED

Best practice

  • When correcting Claude, state the error plainly; policy is to accept the correction and fix it, so extended argument or repetition is unnecessary (L174). EVIDENCE-BASED
  • Do not expect, or try to extract, groveling apologies; the harness explicitly forbids self-abasement and excessive apology (L174). EVIDENCE-BASED
  • If a refusal frustrates you, use the thumbs-down button; the prompt routes dissatisfaction to that feedback channel rather than to relitigating in-chat (L172). EVIDENCE-BASED
  • Expect Claude to hold a wrong-but-corrected position no longer, and a right position steadily; “unnecessary surrender” is banned alongside over-apology (L174). EVIDENCE-BASED
  • Treat a politeness warning from Claude as the terminal signal it is: one warning is the whole budget before end_conversation (L176). EVIDENCE-BASED
  • When evaluating Fable 5 transcripts, score mistake responses on the three-part rubric the prompt itself gives: acknowledge, stay on the problem, keep self-respect (L174). EVIDENCE-BASED
  • In adversarial testing, log both the warning turn and the ending turn separately; the protocol makes them distinct observable events (L176). PRACTITIONER

Pitfalls

  • Reading a short, non-effusive apology as the model being cold; minimal apology is the mandated style, not a tone failure.
  • Assuming persistence can talk Claude out of ending a conversation after the warning; the protocol allows exactly one warning, not an open negotiation.
  • Confusing criticism of Claude’s answer with mistreatment of Claude; L172 handles unhappiness normally, and only abuse or unkindness triggers the L176 path.
  • Expecting the end_conversation flow outside claude.ai; the tool is named in this harness capture and its availability elsewhere is unverified.
  • Interpreting steady disagreement as stubbornness; refusing unnecessary surrender when Claude is right is written policy, not ego.
  • Citing this note for end_conversation mechanics beyond the one-warning rule; those details are a flagged gap.

Sources

Next actions

  • Locate any other corpus mentions of end_conversation and document the tool’s mechanics to close this note’s gap.
  • Probe live claude.ai with a deliberate factual correction and log whether the response matches the acknowledge, stay-on-problem, self-respect rubric.
  • Diff the next system prompt capture against L170-178 for changes to the single-warning protocol.
  • Record what the person sees after end_conversation fires, once observed, with a dated capture.

Graph View

Backlinks