Breakdown · Theo - t3․gg
A Proper Guide to Fable 5: Shipping at the Speed of Psychosis
Fable 5 isn't just a smarter version of the AI models you're used to; it’s a fundamental shift in how software gets shipped. If you're still prompting it like it’s just another chatbot, you’re leaving 90% of its power—and your productivity—on the table.
The New Era of AI-Driven Shipping
The return of Fable 5 marks a turning point for developers who have felt stuck in the 'circles' of previous models like Opus 3.5 or GPT-4. It isn't just an incremental upgrade; it represents a fundamental change in what automated agents can achieve. While previous models might help you write a function, Fable 5 is capable of handling end-to-end implementations, from testing and verifying to breaking down complex tasks into sub-agents. For developers who have struggled to finish long-standing projects, this model offers a way to clear backlogs that have sat stagnant for months.
“In the first day of having it back, I got more work done than I had in the month prior. It just, it was incredible to watch as the model cranked through all of these things that I'd been planning and thinking about and starting, but not finishing.”
The Reasoning Trap: Why 'High' Beats 'Max'
A common mistake new users make is cranking the reasoning effort to 'Max' or 'Extra High,' assuming more thought equals better code. In reality, these settings often lead the model into recursive loops where it over-engineers simple solutions, resulting in bloated code and astronomical costs. The 'High' setting is widely considered the sweet spot—it provides enough intelligence to solve complex problems without the 'psychosis' of over-reasoning. Sticking to the defaults can literally cut your API bill in half while delivering cleaner, more direct implementations.
“X High and Max end up second guessing themselves too much because they run in loops. They'll just go longer and longer with their reasoning and the result is often worse.”
Architecting the Brain with CLAUDE.md
To maximize Fable's potential, you must treat your global configuration files as a glossary for the agent's behavior. By defining a model hierarchy in your CLAUDE.md, you can teach Fable when to delegate tasks to cheaper models like GPT-5.5 (for mechanical work) and when to take the lead on tasks requiring high 'taste'—such as API design or UI copy. This creates a cost-effective ecosystem where the smartest model acts as an orchestrator rather than a brute-force worker for every single line of code.
“Escalating costs less than shipping mediocre work. Judge the output, not the price tag. This both helps me explain to the model what I have in mind when I'm describing tasks and work and the issues I have with things.”
Teaching Skills: The Hybrid Workflow
One of the most powerful ways to use Fable is by teaching it to 'shell out' to other tools. By creating custom skills, you can enable Claude to invoke external CLI tools like Codex for independent code reviews or browser automation. This hybrid approach allows you to use Fable for the high-level logic while leveraging other models for specialized tasks like computer use or massive data analysis. This prevents the primary model from getting bogged down in token-heavy tasks that don't require its full reasoning capabilities.
The Autonomous Sprint: Merging to Staging
The ultimate test of this workflow is the autonomous sprint. By setting a high-level goal and giving the model permission to manage its own work trees and branches, you can essentially leave the agent to clear a month's worth of roadmap while you're away. In the Lakebed project, Fable triaged 16 open PRs, decided which to scrap and which to merge, and executed the entire plan while maintaining a staging environment. The key is keeping a human in the loop for production deployments while letting the agent go 'insane' in staging to prove the code works.
“My whole month. Roadmap that I'd been running around in circles, not shipping, just thinking too much about now is in lakebed. It did the thing.”
The Future is Parallel and Mobile
Shipping doesn't have to happen at a desk anymore. Tools like T3 Code and mobile wrappers allow you to monitor and trigger agent workflows from your phone. By parallelizing work across multiple machines—connecting via SSH or Tailscale—you can have five different threads solving five different bugs simultaneously. This 'multi-agent' reality changes the developer's role from a writer of code to an editor and architect, judging the speed and complexity of the model's solutions to identify architectural weaknesses in the codebase.
Key takeaways
- Never use 'Max' reasoning; 'High' reasoning is more efficient, less prone to loops, and significantly cheaper.
- Use CLAUDE.md to define a hierarchy of 'Intelligence' vs. 'Taste' to help the model choose the right sub-agent for the task.
- Automate your staging environment by giving the model permission to rebase, branch, and merge PRs autonomously.
- Monitor the time it takes an agent to solve a problem; if a fix takes over an hour, it's a signal that your architecture needs human intervention.
- Leverage custom skills to 'shell out' to other models for specialized tasks like computer use or independent reviews to save on token costs.