The Fable 5 Playbook: Capture a Model's Reasoning Before You Get Charged Per Turn

Raj (@askrajgpt) argues that most people rely on a powerful model but few build a system that makes every model powerful. His fix is a document he calls the Fable 5 Playbook, built by asking Claude to write out how it reasons, runs workflows, structures prompts, and formats deliverables. The idea is to extract the working habits of a top-tier model into a reference file, then point your other models -- Opus, Sonnet, Haiku, ChatGPT, or Codex -- at that same file so they reason in a similar way. His framing: the model is rented, but the playbook is yours. He pitches it as something to save before your next Claude Code session, and before access terms or per-turn API pricing change the economics.

The Premise

Raj says a viewer told him the move was to copy the workflows, prompts, and deliverables from a strong model. He took it further and decided to also copy the reasoning, because how the model thinks matters most to him. He then went into Claude and asked it to write out its own playbook: how it thinks, how it reasons, how it handles its different tasks, and how it would upgrade its system prompts to encode that behavior. The output is a single reference document that other models can load so they inherit a similar level of verification, lookup discipline, and confidence scoring. He is clear that this will not produce identical results across models, but it gets you further toward the higher-tier behavior.

The On-Screen Playbook

The video is a single static graphic titled "Fable 5 Playbook" with the subtitle "The four parts that make any model work like your best one." It is split into four labeled panels. The footer reads: "The model is rented. The playbook is yours."

P1 -- Reasoning

• Verify the 2-3 claims that would break the plan, not everything
• Trust order: tool output over memory over guessing
• When surprised, find the cause before retrying
• Say which parts are verified and which are estimates

P2 -- Workflows

• Fan out helper agents in parallel: measure, analyze, verify
• Every step paired with a check that shows real output
• Nothing ships below a 95 confidence score
• Long jobs run detached and resume where they stopped

P3 -- Prompts

• Give each helper agent exact steps and output format
• State what NOT to touch, and demand exact counts
• Repair agents: never delete, log every change, flag conflicts
• Spot-check every number a helper returns

Deliverables

• First sentence answers the question
• Findings ranked, one clear next step at the end
• Numbers verified or labeled estimates, never blended
• It challenges you with a stronger idea, not just agreement

How To Use It

The workflow Raj describes is straightforward. Ask Claude to produce this playbook, save it as a reference document, then have your other models read it at the start of a session so they follow the same reasoning and verification patterns. Because it is a plain document, it is portable across tools -- he names ChatGPT and Codex alongside the Claude model tiers. In the video he offers to DM the document to anyone who comments "Fable."

The description reframes the pitch: capture the model's best habits now so your system keeps them long after access changes. The four things it comes down to are reasoning, workflows, prompts, and deliverables.

Key Takeaways

• The playbook is a reference document created by asking Claude to write out its own reasoning, workflows, prompts, and deliverable habits.
• It is meant to be portable: point Opus, Sonnet, Haiku, ChatGPT, or Codex at the same file so they inherit similar behavior.
• Reasoning rules emphasize verifying only the 2-3 claims that would break a plan, trusting tool output over memory over guessing, and labeling what is verified versus estimated.
• Workflow rules call for fanning out parallel helper agents, pairing every step with a check, shipping nothing below a 95 confidence score, and running long jobs detached so they can resume.
• Prompt rules stress exact steps and output formats, stating what not to touch, demanding exact counts, and spot-checking every number a helper returns.
• Deliverable rules want the first sentence to answer the question, findings ranked with one next step, numbers kept as verified or estimates rather than blended, and the model challenging you with a stronger idea instead of just agreeing.
• Raj's core argument: the model is rented and access can change, so extract its habits into a document you own before pricing or terms shift.