How I Make Opus Think Like Fable (5 easy steps)

Nate Herk · 9:59 · The model isn’t the moat — the process is. How to extract a frontier model’s working discipline into a reusable skill so cheaper models execute at frontier quality.

The thesis

• Karpathy handed Sonnet 3.7 beats a beginner handed Fable 5. What you build around the model (instructions, systems, loops) matters more than the model.
• His test: dynamic workflows where Fable orchestrates all-Fable vs Fable orchestrating Opus/Sonnet sub-agents → about the same results, exponentially different cost.
• “You can’t keep the model’s intelligence, but you can keep its process.”

Frontier model = teacher, not workhorse

• Use Fable/Frontier to improve your setups and skills, i.e. treat it like a senior engineer packaging everything it knows for the junior cohort that replaces it.
• Lessons picked out of the leaked Claude Fable 5 system prompt:
  • Partial recognition from training ≠ current knowledge — verify memory claims; a prompt implying a file exists doesn’t mean one does (check it exists).
  • Ambiguous query: answer first, then ask — one question max.
  • Acknowledge what went wrong, stay on problem, maintain self-respect.
  • Effort calibration in the prompt itself: 1 iteration for signal facts, 3–5 for medium tasks, 5–10 for deep research/comparisons.

Effort level > model choice

• Release-blog chart: Fable 5 at low effort ≈ Opus 4.8 at high (quality/cost curve). Same task, wildly different price.
• Higher isn’t better: max-effort runs go longer, cost more, overthink, second-guess, and produce something worse than high effort.
• Match effort to task the way you match model to task.

The extract-the-method move (“Fable mode”)

• Take any Fable deliverable you loved but couldn’t explain why → have it (or Opus) analyze the session: What did you think to get here? How did you prove it worked?
• Distill that into an installable skill. Prompt shape:

  “Write a complete installable skill file that makes Opus 4.8 operate with your judgment, your planning, verification, and reasoning habits” — call it fable-mode.

• His skill enforces five gates: scoping before work → evidence before reasoning → reasoning adversarially → verifying before declaring done → calibrating/reporting. Any model can run it (GPT, open source too).
• Key scoping nuance: planning ≠ just listing steps. The frontier-model habit worth copying is playing devil’s advocate — enumerate everything that could go wrong and every unknown before executing. That’s why his “Fable plans, Sonnet executes” workflow matches all-Fable quality: Sonnet workers report back, Fable keeps redesigning next steps.

Model routing table

• Give your agent a literal table of available models scored on: cost, intelligence (code review, comprehension), taste (creativity, UI/UX) — plus any categories your work needs. Codex/open-source models can be delegates too.
• Real test he ran: Opus orchestrator + Haiku scout workers vs Sonnet/Opus workers → Haiku delegation was ~3× cheaper, identical result.
• This is the unit-economics skill that separates people getting far more output per dollar.

Why this matters now

Fable’s subscription availability has been on/off; Anthropic says it returns, timing unknown — but “we don’t own anything.” What you can own: processes, systems, methodologies — plus hardware and local models. Build the discipline into artifacts you control, not into whichever model currently tops the leaderboard.