So I was using Fable wrong...
Tuesday, Sep 22 · 42 min · about 10 min to read
Theo-brain watched all 42 min
Theo shares how he prompts Claude Fable 5.1 and why most old agent setups are out of date. The main lesson: tell the agent exactly where the job ends, have it check its own work, and start fresh with default settings, adding small fixes only when a real problem appears.
Skim itWatch the stretches where Theo shows his real prompts and explains his mindset; the run through the guide itself is covered here.
Tap any time to jump the video there.
Key takeaways
Every prompt should say where the job ends. Examples: babysit the pull request (a proposed code change) until all checks pass, wait for passing checks and then merge, or just file it and report back. Theo says this is the key thing to copy from how he prompts.
sure you guys take out of how I'm prompting is the end states. I make it clear to the model where I want it to be done. Sometimes I'll say I want you to babysit it until everything is green. Sometimes I'll say I want you to wait till it's green, then merge it.
Theo thinks your setup and your beliefs about agents are probably out of date. He suggests starting over: install Claude Code fresh with as little setup as possible, then add small fixes only when a real problem appears.
here are simple. The tools are good now. Any customization or careful configuration you've done to work around the flaws are probably out of date. And a lot of your own mental model around this stuff is also out of date. Reset it. Pretend you're on a brand new machine, delete everything, install Claude Code from scratch, set it up with as little as possible. I bet you'll be
Leave the effort (reasoning) level on high. On low and medium, Fable fails often enough that redoing the task costs more tokens than running it on high once. On simple tasks, high doesn't waste much.
With low and medium, I find that it's likely enough to fail on those reasoning levels that I try to avoid them just because it ends up being more tokens if you have it do the task on low and it fails and then you redo it on medium or high. I'd rather just have it run the
Fable 5.1 often goes long stretches, like 150 tool calls, without an update. Theo doesn't mind because he only wants to see the result when it's done. If you want updates, just ask for them in the prompt.
this, too. I can't tell you how many times I had Fable just go off and do like 150 tool calls without giving me an update. And you know what? I don't care. I don't want that info. I just want to see it when it's done. That's how I
But if you do, and you want to get these updates, there's a really simple solution. Ask for them. I do
Delete the behavior and formatting rules in your Claude MD or Agent MD files (standing instruction files the agent reads). See how it works without them, then add back only what you need. Theo has none in his and says it has been pretty nice to work with.
speaking, any behavioral instructions that you have in your Claude MD are worth deleting and reconsidering. See how it feels to not have them at all, and then adjust over time as you need.
them for now. I have nothing like this in my Agent MD or Claude MD, and it's been pretty nice to work with. Also of
Two skills here are new to you.
A "babysit PR" skill
- What it is
- A saved instruction set (skill) that kicks in when Theo tells the agent to watch a pull request. It keeps checking for new review comments and failed checks, verifies bot findings before acting, fixes real problems and keeps the change up to date with the main code.
- Why Theo cares
- When he comes back to the thread, the change is ready to merge. He added "don't scope creep" and "don't post filler comments" after some models added new work or left needless comments.
- How to try it
- Ask your agent to write a skill that monitors a pull request for new comments and checks, and fixes only real findings.
- Include rules: verify each bot finding against the code, don't add new features, don't post filler comments.
- End prompts with "babysit it until everything is green".
skill. This is a skill that gets invoked when I tell the model to watch or babysit a pull request. The point of this skill is to keep the model checking for updates, for new comments, and make sure by the time I go check the thread, it's good to go. The skill's comments
Have Fable call Codex or Astra to check its work New
- What it is
- While Fable does the coding, it hands review or testing to another model. Theo uses OpenAI's Codex for computer use (clicking around apps to verify results) and Astra for review.
- Why Theo cares
- He finds Fable the much better code writer, but Astra a slightly better reviewer, and Claude weaker at computer use than Codex, especially on macOS. Fable then fixes problems and retests before reporting back.
- How to try it
- If you have even a cheaper Codex subscription, set it up on your computer.
- Tell Fable: "Can you have Astra verify this?"
- Let it update and retest until it is happy before it returns.
harness has access to. I do still find that Claude isn't the best at computer use compared to Codex, especially on macOS. So, if you have a Codex sub, even like a cheaper one, set up and on your computer, you can have Fable call Codex to do the computer use stuff to verify
recommend and how they review things, I found Astra to be a slightly better reviewer than Fable. But, obviously Fable's a much better writer of code than Astra. As such, I often will have Fable just go ask Astra to give it feedback on some changes or to go test the changes it made in order to verify
Anti-scope-creep instruction from Anthropic New
- What it is
- A line from Anthropic's guide: if the agent finds existing bugs or performance issues the task didn't mention, it should not fix them in this change but list them as follow-ups in its summary.
- Why Theo cares
- Anthropic says it cuts unrequested additions sharply with no drop in task success. Theo hasn't seen much scope creep himself, but says you can just paste this in if you have the problem.
- How to try it
- Add this instruction to your Agent MD or system prompt only if your agent keeps touching unrelated code.
testing, you find pre-existing bugs, performance concerns, or behaviors the task doesn't mention, don't fix, optimize, or extend it in this change unless the requested behavior cannot work without it. Report it as a follow-up in your summary. This is
Advice
If context limits, compaction or settings confuse you, don't change the defaults. Theo's only Claude Code changes are full screen and routing through his proxy. He also turned off Claude Code's memory because it wasn't useful to him.
recently. The best thing you can do if these things are confusing or concerning to you is don't change the defaults. The only settings I have changed in my Claude code config on my machine are I have it full screen by default, and I have it route through my proxy layer. That is it. The defaults in Claude code
at Anthropic. Okay, memory's a fair point here actually. I did turn off memory. I do not like the memory in Claude code. I'm sure it is useful in some places in some ways. It is not useful for me. There was actually just
If Fable's writing feels flowery, tell it "please remove all mannered prose". Mannered prose means showy metaphors in place of plain statements. You don't need Anthropic's long definition.
perform. If you don't want to write all this in your system prompt, or use the admittedly not good examples, just say, please remove all mannered prose. The
Only try his merge-and-deploy hands-off workflows if you have good staging and testing in place. Theo calls them dangerous on big, important codebases.
next example is not meant to be a thing that you do in your day-to-day work on like a real big important code base cuz it's a dangerous. If you have really good systems for staging environments and QA and testing before things go out, maybe play around with these types of workflows cuz they are super super fun.
Hot takes
Fable isn't worth it at full API prices. Only use it through a subscription.
His reasoning: Paying full API price seems insane to him. Astra isn't a fix either, because it can get stuck in loops and spend far more than it should.
up front with this. I don't think Fable is worth the money if you're not getting it subsidized through a subscription. Paying the full API price just kind of sounds insane to me. I don't think Astra is much better here, sadly, because Astra can often get stuck in a loop and end up spending way more than it should.
The "dumb zone" doesn't matter much. That's the fear that a long conversation makes the model worse. You don't need to watch how much context you've used.
His reasoning: Context limits are chosen numbers. Anthropic set one million tokens where it sees real degradation begin, and models are now good at compacting (summarizing) long runs.
context to Fable. They pick a number based on the behaviors they see. And Anthropic picked one million tokens of context because that is the point at which they think you might start to see enough degradation that you shouldn't go further. The dumb zone doesn't matter that much. The models have gotten pretty
The "accept edits" and "supervised" permission modes are relics. Use auto mode if you're careful and full access if you're not.
His reasoning: He says the permission systems and auto-approval modes are very good now.
don't have to set that up yourself. The permission systems and the auto approval modes are very good now, and you definitely shouldn't use accept edits or supervised. We recently added
supervise and auto accept edits are tools from the past, from like 2024 even at like newest. Use auto if you're careful and use full if you're not, period. The points I'm trying to make
Walking the model step by step through fixing, testing and filing a change it could handle alone wastes your brain. It's like writing a redundant type check.
His reasoning: If the computer can do the verifying, using your own attention for it frees nothing up for more important work.
work. If the model can root cause a bug, fix it, verify that it's fixed, film a video showing the results, put up a pull request, monitor it to address all the review comments as they come in, and then tell you when it's done, if you're going through those steps with the model, you are the same person as the one who would write the type check after the type check in a TypeScript function.
How Theo thought about this
Astra wrote a risky rewrite of his Lakebed runtime. Rather than merge it, Theo asked Fable for an honest review, then asked whether it could debug problems in staging (a test copy of the live system). Fable said it couldn't catch wrong query results there, so he had it build the checks it needed first.
Instead of just blindly merging Astra's crazy work here, I asked Fable what its thoughts were. I took a screenshot of the claims Astra had alongside the PR and said, "Is this ready to merge? Give
change. It said it could partly debug crashes, latency, and resource problems in staging, but it couldn't confidently debug wrong query results, and that's the scary class of bug. This is a really
His role has shifted. He spends less time choosing between options the model offers and more time giving it what it needs to choose well. Sometimes he gives it two paths: come back if unhappy, otherwise keep going.
for me is I'm I'm spending less time trying to choose from the options the model gave me and more time trying to give the model everything it needs to make good choices. It gave a summary with the
paths that we can take. If it isn't happy, then it can tell me that and we can talk about it. If it is happy, then it will start working. I don't have to come back. I'm trying to minimize the amount of times I go back to a thread until the work is done. And here I gave
Sign in to mark what you've watched and keep notes.
Have a question about this one? Ask about this video →
Theo Daily is unofficial and not affiliated with Theo or t3.gg. Written by Claude Opus 5.5 from each video's transcript.
