Claude Fable 5 Prompting: What Actually Changed
I've been running Fable 5 daily since launch, starting with six measured experiments in its first 36 hours (all documented in The Ultimate Guide to Claude Fable 5). Prompting it is not like prompting Opus 4.8 or GPT-5.5. Five shifts matter.
Searching for Fable's system prompt? I captured the real one Claude Code runs. It's in the last section.
1. Set Effort, Not Thinking
You can't turn thinking off. Pipelines that set thinking: disabled get a 400 error. What you control now is the effort dial, and I measured it across 5 rounds of experiments: every effort level solved the same problems with the same method. The extra seconds bought re-verification and caveats, not correctness.
Default to high. Save max for rare, genuinely hard calls: on easy questions it triples response time and swings wildly between runs.
Temperature is gone too. If your eval suite sweeps sampling settings, those configs fail on Fable.
2. Delete the Scaffolding
Step-by-step recipes made older models reliable. They make Fable act like an older model.
When I let Fable read the 166 instruction files my agents follow, it flagged the guardrails I had written for weaker models: double-check rituals, "restate the plan before acting," recipes for judgment calls. Deleting them made outputs better, not worse.
The test: would you give this instruction to a strong new hire on day one? "Never publish without my approval" passes. "Before every paragraph, re-read the style guide" does not.
3. Hand It Objectives, Not Tasks
Fable can run long tasks end to end: it plans, delegates to subagents, and verifies its own results. The prompt shape that unlocks this:
/goal [outcome] + [evidence required] + [boundaries] + [stop condition]For example:
/goal Synthesize the attached 40 interview notes into ranked opportunity
areas. Evidence: every opportunity cites at least 3 interviews by ID.
Boundary: don't invent segments not in the data. Stop when I have a
ranked table plus the 5 verbatim quotes I'd put in front of leadership.The difference from a normal prompt: it defines done. That is what lets the model run for hours without you. The full pattern is my Intent Engineering Framework: lead agents the way good orgs lead empowered teams, with context, not control:
4. Run the Audit Before Real Work
Fable reads instructions critically, not obediently. So the first prompt to run is not a task:
Read your own instruction files (CLAUDE.md, skills, rules, memory files) end to end.
1. Where do they contradict each other? Quote both sides.
2. Which rules exist to manage a weaker model: guardrails for failure
modes you don't have, recipes for things you no longer need spelled
out, hardcoded facts that have drifted? List them with file:line.
3. Which rules teach by bad example: documents that violate the
patterns they prescribe?
4. What would you delete? What would you keep exactly as is, and why?
Don't fix anything yet. Report first. I decide what gets cut.Mine found a hardcoded date telling every session the wrong day, a rule documented with the exact pattern it bans, and calibration numbers two months stale. The audit is the model's judgment. The deletions are yours.
5. When It Refuses, Check the Routing
Fable ships with classifiers that reroute about 5% of sessions to Opus 4.8, with a notification. Claude Code's Boris Cherny:
"We know the classifiers are trigger-happy, and are working on improving it."
If your work touches security, healthcare, or biotech, expect occasional routing. You did nothing wrong, and the notification tells you which model actually answered.
The Actual Fable 5 System Prompt
I captured the system prompt Claude Code runs on every model, including Fable 5. The prompts and the capture method are in my public experiments repo.
The finding: frontier models get a stripped prompt. Opus 4.8 and Fable 5 run on a prompt cut by about 70% (on Opus: 2,686 words down to 830 with memory on, 514 with memory off). Run Sonnet 5 or Haiku 4.5 and it's still the full 2,094 words, every rule intact: "don't add abstractions," "don't add error handling," "don't write comments." Anthropic's stated reason: the smarter the model, the fewer instructions it needs.
That is shift 2, confirmed by the vendor. If your CLAUDE.md keeps growing, you're writing for a weaker model than the one you're running.
For your own files, the same split applies: keep your constraints (voice, approval gates, non-negotiables), date-stamp your calibrations, delete your scaffolding.
Go Deeper
The Ultimate Guide to Claude Fable 5: the six experiments behind this page. The paid half covers the full migration workflow and /goal patterns.