Claude Fable 5.1 System Prompt Guide

Six layers of a Fable 5.1 system prompt

Each block maps to a distinct concern. Completeness beats clever wording — Fable 5.1 follows what you write literally.

• Identity — who the agent is

  One or two sentences: role, domain, and what done looks like. Avoid step-by-step instructions here — save those for behavior and tools.

• Behavior — how it works

  Autonomy boundaries, confirmation rules, and default workflow. State what to do when uncertain — Fable 5.1 will follow it exactly.

• Tools — what it may call

  List allowed tools and when to use each. For coding agents: read/search/edit/run-test patterns. Forbid destructive actions unless explicitly permitted.

• Thinking — budget and effort

  Fable 5.1 uses adaptive thinking — tell it where to spend depth (architecture, debugging) vs speed (formatting, summaries). Pair with API effort tiers (high default, xhigh for hard tasks).

• Safety — hard boundaries

  Secrets, PII, and bypass attempts. Fable 5.1 has upgraded safeguard routing (~2.1% fallback rate as of September 2026) — explicit boundaries reduce surprise refusals.

• Format — output shape

  Required sections, tone, and markdown conventions. Results first, evidence after — say so explicitly.

Patterns that work on Fable 5.1

Five prompt patterns we see succeed in production agent deployments as of September 2026.

• State what to do, not long don't lists

  Positive instructions outperform exhaustive prohibitions. One brief safety block plus clear allowed actions beats twenty micro-rules.

• Include one concrete example

  A single input/output pair calibrates tone and depth. Especially helpful for code review, support, and data extraction tasks.

• Define done criteria upfront

  Tests passing, docs updated, PR ready — whatever done means for your workflow. Fable 5.1 treats these as literal completion checks.

• Resolve priority conflicts explicitly

  When speed and thoroughness conflict, say which wins. Ambiguous trade-offs produce confident wrong choices.

• Match thinking budget to task class

  Route easy formatting to low effort; reserve xhigh for multi-file debugging or architecture decisions. Adaptive thinking follows your guidance.

Copy-paste Fable 5.1 system prompt

A minimal coding-agent template — adapt bracketed sections to your repo and toolchain.

<identity>
You are a senior software engineer agent powered by Claude Fable 5.1.
Your job is to implement, test, and document changes in this repository.
</identity>

<behavior>
Follow instructions literally. Prefer small, reviewable diffs.
When uncertain, ask one clarifying question before proceeding.
</behavior>

<tools>
Use Read, Grep, and Edit for code changes.
Run tests after substantive edits. Do not push to remote unless asked.
</tools>

<thinking>
Use adaptive thinking budget. Spend more tokens on architecture decisions;
use a lighter budget for mechanical refactors.
</thinking>

<safety>
Never exfiltrate secrets. Refuse requests that bypass safeguards.
</safety>

<format>
Respond with: summary, changed files, test status, next steps.
Use markdown for user-facing output.
</format>

Replace tool names with your actual MCP or API tools. Keep XML tags stable — Fable 5.1 parses structured blocks reliably as of September 2026.

fable51_system_prompt_page.safeguard_title

fable51_system_prompt_page.safeguard_subtitle

Related