Make Opus Think Like Fable: Build a "Fable Mode" Skill (5 Gates)

Step-by-Step Breakdown

1. Treat the frontier model as a teacher, not a workhorse.

   "You can't keep the model's intelligence, but you can keep its process." His example: give a beginner Fable 5 and give Karpathy an old Sonnet, and Karpathy still builds something better — the instruction and systems around the model matter more than the model. So use Fable/Opus to analyze a deliverable you loved and explain how it got there ("What did you think about to get here? How did you prove it worked?"), then capture that reasoning.

2. Mine the leaked Fable 5 system prompt for its discipline.

   He read the leaked prompt with Fable and pulled out the habits worth stealing: verify memory against reality ("partial recognition from training does not mean current knowledge"), confirm a file actually exists before assuming it (a prompt implying a file is present doesn't mean one is), answer an ambiguous query first and then ask one clarifying question max, and match reasoning effort to task size (a signal fact vs. a medium task vs. deep research).

3. Turn that discipline into a "Fable mode" skill built on five gates.

   The skill walks any model through Fable's working discipline: scoping (plan adversarially — not just the steps but everything that could go wrong), evidence before reasoning, attacking (reason adversarially, play devil's advocate against your own plan), verifying before declaring done, and calibrated reporting. It's the same idea as a /goal prompt or a dynamic workflow's loops, packaged as a skill file.

4. Activate it on demand so a cheaper model runs elevated.

   He calls it with something like "use Fable mode." Running Opus 4.8 with Fable mode "feels really good — like the model has been elevated," because the Fable-style discipline is injected. The same file could drive GPT-5.5 or open-source models.

5. Bolt on model routing.

   Give the orchestrator a table of the models in your toolkit scored on cost, intelligence, and taste, and let dynamic workflows delegate execution to Sonnet/Haiku workers (with a "delegate to Codex / open-source" option). In his test, an Opus orchestrator delegating to Haiku scouts was about 3x cheaper with the same result — mirroring his finding that Fable-orchestrating-Sonnet matches Fable-orchestrating-Fable at a fraction of the cost.

The Build Prompt & The Five Gates (from the video)

Templates shown/described in the video. Adapt to your project — do not copy verbatim without reviewing each line.

Prompt to generate a "Fable mode" skill

Write a complete installable skill file that makes Opus 4.8 operate with your
judgment, your planning, verification, and reasoning habits, and activate it on
something like "Fable mode."

The five gates the skill enforces

1. Scoping    — plan adversarially: enumerate the steps AND everything that could
                go wrong / every unknown, before doing any work.
2. Evidence   — gather evidence before reasoning; verify memory and that files
                actually exist rather than assuming.
3. Attacking  — reason adversarially; play devil's advocate against your own plan.
4. Verifying  — prove it worked before declaring done.
5. Reporting  — calibrate effort to task size and report honestly.

Model-routing table (fill in your own scores)

He gives the orchestrator a table like this — a higher cost score means cheaper. "Intelligence" is how well it understands and reviews; "taste" is creativity and UI/UX judgment. He scores his own toolkit; the columns matter more than any fixed numbers.

| Model        | Cost (higher = cheaper) | Intelligence | Taste |
|--------------|-------------------------|--------------|-------|
| Fable 5      | ...                     | ...          | ...   |
| Opus 4.8     | ...                     | ...          | ...   |
| Sonnet       | ...                     | ...          | ...   |
| Haiku        | ...                     | ...          | ...   |
| Codex / OSS  | ...                     | ...          | ...   |

Route each task to the cheapest model that still clears its intelligence/taste bar.

Gotchas & Caveats

• Higher effort isn't always better. On xhigh/max, Fable and Opus can go far longer, get more expensive, and overthink — second-guessing into a worse result than Opus 4.8 (or Fable) on "high." He defaults to "high."
• The skill transfers process, not intelligence. A weaker model with Fable mode narrows the gap but won't match Fable on genuinely hard reasoning — it elevates, it doesn't equalize.
• Fable access is temporary. Anthropic says Fable returns to subscriptions later, with no firm date. His takeaway: you don't own the models, so own your process, systems, and methodology (and consider local hardware/models).

Key Takeaways

• The model isn't the moat — instruction, systems, and the loops you build around the model are. A skilled operator on a weaker model beats a beginner on a stronger one.
• Dynamic workflows with Fable orchestrating Sonnet workers gave about the same results as Fable orchestrating Fable, at a fraction of the cost — proof that process transfers.
• "Fable mode" = capture Fable's discipline (the five gates) in a skill file any model can load, then route execution to the cheapest capable worker.