Django co-creator Simon Willison tested a prompting approach in Claude Code that lets the frontier Fable 5 model itself decide when to spin up cheaper subagents for coding work, keeping only the heavier judgment tasks in the main loop.
The most interesting Fable tip I've heard so far is to let the model use its own judgement as much as possible
I told it "For all coding tasks use your judgement to decide an appropriate lower power model and run that in a subagent" and it seems to be saving a lot of tokens
The persistent instruction tells Fable to evaluate every coding task, spawn a subagent with an appropriate lower-power model for mechanical edits, and route substantive implementation or review back to itself before any commit.
Willison reports the Fable usage limit is now shrinking less quickly on routine tasks, though the post supplies no exact token counts, percentages, or duration of the trial.
Users praise letting models judge task complexity to route work to cheaper submodels or subagents because it delivers massive token and cost savings without losing performance.
No Digg Deeper questions have been answered for this story yet.
What if the model is the router?
I think people underestimate the ability of frontier models now, but especially in the near future, to delegate work on their own as needed to dumber, cheaper models. The future might be start with a smart AI planner, let it delegate on its own.
The most interesting Fable tip I've heard so far is to let the model use its own judgement as much as possible
I told it "For all coding tasks use your judgement to decide an appropriate lower power model and run that in a subagent" and it seems to be saving a lot of tokens
More notes on my blog: https://simonwillison.net/2026/Jul/3/judgement/
The most interesting Fable tip I've heard so far is to let the model use its own judgement as much as possible
I told it "For all coding tasks use your judgement to decide an appropriate lower power model and run that in a subagent" and it seems to be saving a lot of tokens
@simonw Yeah I really like the subagent implementation piece of the superpowers plugin. It does a really nice job delegating to cheaper models dynamically. So I start a plan with Fable and let it do its thing. You have to try that plug-in if you've not
@simonw I built a tool around the same idea! I want fable to delegate workflows to Codex (which has much higher usage limits)
https://github.com/YuanpingSong/ultracodex
@simonw Its amazing, you can also say "look for our previous chats and infer what I would do for this specific problem, use subagent to get info you need", it infer pretty well your pattern of intent and follow suit.
@simonw
@simonw Yep. The better the underlying reasoning, the less scaffolding they need. Also applies to the personal system prompt/instructions. Beyond a certain point, the personal ‘preferences’ tend to reduce performance by adding noise and unnecessary constraint.
@simonw This is the exact mechanic behind a skill I built this week, the plan gets decided once, internally, never printed as a deliberation. Benchmarked it against a viral 'minimal code' skill on its own suite: https://github.com/anshaneja5/scalpel
@singular_prism @simonw Yeap. Just tell Claude to set it up. I give Claude access to all the other CLIs. https://github.com/Jdjohnson/skills/tree/main/downloads/cli-skills
@simonw The most awesome Fable thing to me is free usage. Not sure how I did it, but I glitched the system. Fable medium has been working for 20 hours now without a single % point increase in Fable usage.
I told it no agents, just in case.
Some kind of 2 day old session bug
@simonw Interesting. I’ll give that a try just to test it
@simonw Tried this exact pattern last week — let Claude self-select model tier per subtask instead of hardcoding Opus everywhere. Cut my API bill by ~30% on a multi-step pipeline. The model is genuinely good at knowing when it doesn't need the expensive tier.
@simonw Can we ask it to use codex and openai as well
I frequently say at the end of a prompt something like
“Of course I’d prefer that rather than attempt ‘exactly what the user says’, you recognize that I am less intelligent and don’t read the code and whatever insight I have is best extracted and totally subsumed by your judgment in the end thanks genius <3”
@_markfenner @simonw Yeah, same. Wonder if we can also tell it to use other cheaper models during its task and just act as the orchestrator.
@simonw My official position is that anything that would theoretically work on a human will probably more likely work on Fable 5
Doesn’t matter if it’s conscious. It behaves in reaction to you
@simonw we ask the model to make decisions constantly until it makes one we didn't expect, then we yank control back. letting it actually pick the subagent is the harder trust exercise than it sounds.
@simonw Been doing the same, even giving it Codex to delegate to…works great!
@simonw I've adopted having fable captain the work. Fable requests scans, evals, double-checks, authors a plan, and then spins up minions for the impl. It uses ultracode fan-outs wisely.
@simonw you're trying this in claude code?