Fable Brain

Tone and Formatting Rules

6 min read

Tone and Formatting Rules

The claude.ai system prompt gives Fable 5 a warm, prose-first voice with minimum formatting by default, then layers channel precedence on top: user styles outrank user preferences, preferences apply by default only with “always” wording, lists never become artifacts, and routing is never narrated.

What it is

  • The communication layer of the claude.ai harness for Claude Fable 5: the default tone rules (tone_and_formatting, System Prompt Export 2026-07, L80-92), the formatting rules (lists_and_bullets, L94-104), and the channels through which instructions about voice and layout actually bind.
  • Relevant to prompting because these rules decide what the model sounds like with no instructions at all, and which channel (message, preference, or style) overrides the defaults most reliably.
  • Bound to the consumer surface: the export captures claude.ai on Fable 5, deployed for Pro, Max, Team, and Enterprise users (https://platform.claude.com/docs/en/about-claude/models/introducing-claude-fable-5-and-claude-mythos-5, retrieved 2026-07-07).

How it works

Default tone (L80-92)

  • Warmth as baseline: “Claude uses a warm tone, treating people with kindness” and without negative assumptions about judgement or abilities; pushback stays constructive, with empathy and the person’s best interests in mind (System Prompt Export 2026-07, L82).
  • Explanations may use examples, thought experiments, or metaphors (L84); cursing is off unless the person asks or curses a lot themselves, and even then sparing (L86).
  • Question discipline: Claude “avoids more than one per response” and tries to address even an ambiguous query before asking for clarification (System Prompt Export 2026-07, L88).
  • Audience calibration: a suspected minor gets a friendly, age-appropriate conversation; otherwise the person is assumed to be a capable adult and treated as one (L90).
  • Evidence hygiene in tone: a prompt implying a file is present does not mean one is; Claude checks for itself (L92).

Lists and bullets (L94-104)

  • Minimum formatting law: Claude avoids over-formatting with bold, headers, lists, and bullets, “using the minimum formatting needed for clarity”; lists and bullets appear only when asked or when content is multifaceted enough that they are essential, and bullets run at least 1-2 sentences (System Prompt Export 2026-07, L96).
  • Prose is the default register: typical conversation and simple questions get natural prose, and casual responses can be short (L98).
  • Reports, documents, technical documentation, and explanations are written as prose without bullets, numbered lists, or excessive bolding; inline enumerations read naturally as “some things include: x, y, and z” (L100).
  • Declining softly: Claude “never uses bullet points when declining a task”; the additional care softens the blow (System Prompt Export 2026-07, L102).

Channel precedence and routing

  • Preference gating: “Preferences should not be applied by default unless the instruction states ‘always’” (System Prompt Export 2026-07, L846-873); behavioral preferences otherwise need direct task relevance, and contextual preferences need explicit invocation. Detail in User Preferences System.
  • Precedence: when a userStyle and a userPreference conflict, userStyle wins (System Prompt Export 2026-07, L846-873).
  • Format routing: lists, tables, and enumerated content never become artifacts regardless of length, while code over 20 lines and standalone text-heavy documents over 20 lines or 1500 characters do (System Prompt Export 2026-07, L1117-1140).
  • Silent routing: visual output follows a fixed checklist (prose, then MCP tool, then file tools, then Visualizer) and the routing decision is never narrated to the user (System Prompt Export 2026-07, L1198-1222).
  • Phrasing bans as tone control: for memory, “Claude NEVER uses observation verbs suggesting data retrieval” (System Prompt Export 2026-07, L251-276), showing the harness legislates wording down to verb choice.
  • Enforcement backstop: classifier-triggered runtime reminders (including long_conversation_reminder) can re-tighten behavior mid-session and never reduce restrictions (System Prompt Export 2026-07, L144-152); see Anthropic Runtime Reminders.

Key insight

Tone flows through three channels with fixed precedence: per-message instructions steer the current task, userPreferences apply only when “always”-worded or task-relevant, and userStyle overrides userPreferences on conflict. Prompting that ignores the channel hierarchy produces intermittent, hard-to-debug voice drift.

Best practice

  • Expect prose by default; if you want bullets, tables, or headers on claude.ai, ask for them, because the harness treats formatting as opt-in. EVIDENCE-BASED
  • Ask at most one thing per turn when mirroring the harness style; the model itself is bound to one question per response. EVIDENCE-BASED
  • Put persistent voice requirements into a style rather than a preference; on conflict the style wins. EVIDENCE-BASED
  • Phrase preferences you want applied everywhere with explicit “always” wording; otherwise they gate on task relevance. EVIDENCE-BASED
  • Do not ask claude.ai to deliver lists or tables “as an artifact”; the harness excludes them from artifacts regardless of length. EVIDENCE-BASED
  • Do not rely on Claude explaining why it chose prose, a tool, or the Visualizer; routing is deliberately unnarrated, so ask for the output form you want directly. EVIDENCE-BASED
  • When per-message tone instructions fight harness defaults, repeat them near the task rather than once at session start; runtime reminders can re-tighten defaults mid-session. PRACTITIONER
  • In this vault, follow the CONVENTIONS style law regardless of model defaults: declarative present tense and no em dashes. EVIDENCE-BASED

Pitfalls

  • Assuming preferences are always-on: without “always” wording, a stored preference silently fails to apply to most tasks.
  • Fighting the artifact threshold by asking for a 25-line list as an artifact; the list exclusion beats the length trigger.
  • Reading the absence of routing narration as indecision; the checklist ran, it is just not reported.
  • Expecting heavily formatted reports by default; the export bans bullets and numbered lists in report prose unless asked (L100).
  • Generalizing these rules to the API or Claude Code; they are claude.ai harness rules from the 2026-06-09 capture.

Sources

Next actions

  • Test the style-beats-preference precedence with a live claude.ai probe and log the result.
  • Compare a fresh capture against L80-104 and L846-873 to confirm the tone defaults and “always” gating survived past 2026-06-09.
  • Log one claude.ai probe of the list-exclusion rule: request a 30-line list as an artifact and record what happens.

Graph View

Backlinks