Getting the most out of Opus 5.5
Thursday, Sep 24 · 29 min · about 9 min to read
Theo-brain watched all 29 min
Theo shares how he prompts Opus 5.5 for long agent runs. The core lesson: tell the agent exactly what 'done' looks like, pre-approve what it will need, and give it a clear reason to stop and ask. Leave reasoning on high or xhigh, and never use max.
Skim itThe notes cover the advice; watch the live prompt walkthrough and the max-reasoning rant if you want to see how Theo words his prompts.
Tap any time to jump the video there.
Key takeaways
Define the finish line. Top models from Anthropic and OpenAI keep working until they reach the 'done' you describe, for example: build these three things, show a screenshot, then open a PR (a pull request, a proposed code change).
screenshot. Generally speaking, these higher-tier models from Anthropic and OpenAI work really, really well if you tell them what done is, because they'll keep going until they get there. The
Even a good prompt was not enough. Opus finished only phase zero of an eight-phase plan and stopped, so Theo followed up: hold every question until phase seven is done and don't ask permission between phases.
And it did phase zero. And then stopped. I want you to hold off on all questions until the very end when you get through to phase seven in completion. I don't want you to stop after every phase to ask for permission
In Theo's own benchmark, max reasoning versus xhigh used over 10x the tokens and pushed average answers to 50 seconds, with the slowest 20x slower. The gain was one more correct answer, 78% to 79%.
a 10x. Much scarier, it bumped the average duration to 50 seconds. So the average max run took longer than the slowest X high run. And the slowest run was 20 times slower at 600 seconds. You know what the best part is? All it got out of that was one additional correct answer. It went from 78% to 79%. It cost
Adding instructions while the agent is already working (steering) now works. Older models dropped the earlier tasks when you sent a follow-up; Opus 5.5 treats it as extra direction, so Theo no longer restarts threads.
how it helped the context. They have now been trained better about this to take these interruption messages not as a reason to stop previous work or treat it as done, but as what they are, steering, additional things to help the model go
With no design direction, Opus 5.5 falls back on a few default styles. Saying 'avoid generic looks' just swaps one default for another; a list of specific things not to use works much better.
reason this matters with Opus 5.5 is that with no design direction, Opus 5.5 will fall back on a few default styles. A general instruction like avoid generic looks will mostly just swap one default for another. A list of specific patterns will work much better. For example,
Three skills here are new to you.
The three-part prompt: goal, done, and an out New
- What it is
- Every long task prompt states what you want, what counts as done, and when the agent should stop and ask you.
- Why Theo cares
- Models are trained so hard not to give up that when they have a question they answer it themselves and push on. An explicit 'ask me if X' line makes them willing to stop when genuinely stuck.
- How to try it
- State the goal, then 'Done means…' with checkable conditions such as tests passing or a screenshot.
- End with one line on when to stop and ask, e.g. only if a test fails for a reason it can't explain.
explain. This is the perfect simple one, two, three. What you want, what means done, and the out if it needs an out if it does get confused or does something
silly, but these models have been RL'd to hell and back to not give up on tasks. So, if the model asks a question, the initial thing it's been RL'd to do is to try and answer the question itself. So, if it did have
A 'when to stop' rule in CLAUDE.md New
- What it is
- CLAUDE.md is the instruction file Claude Code reads in every session. Anthropic suggests a rule: keep going when no input is needed, and stop only when blocked or before anything destructive.
- Why Theo cares
- Opus 5.5 often pauses to report or offer choices on long runs. Theo wants to fire off a thread and come back when it's finished, so he pasted this rule into his own CLAUDE.md on camera.
- How to try it
- Add: keep going when a step doesn't need me; put status notes with your next action.
- Add: stop only if you can't continue, or before deleting data, force pushing, or changing anything outside this repo.
- If a run still asks 'want me to continue?', just reply 'continue'.
as well. When a step doesn't need my input, keep going. Put status notes in the same message as your next action. Stop and ask only when you can't continue without me or before anything destructive, deleting data, force pushing, or changing anything outside of
Ask 'What are the risks of merging this today?' New
- What it is
- Before reviewing a large agent-made change, ask the agent how risky it is and what the worst outcome of merging it now would be.
- Why Theo cares
- It is his most common prompt. Long runs produce changes he doesn't want to read line by line, and the answer tells him whether reading them is worth it.
- How to try it
- When an agent hands you a finished PR, ask what the risks of merging it today are.
- Also ask: 'What's the worst thing that would happen if we merge this right now?'
Another one I ask a lot, and I I'm amazed at how helpful this has been. What are the risks of merging this code today? That's probably my most common prompt. I'll often have a PR up from an
Have a different lab's model review the code
- What it is
- Theo codes with Claude-style models but has OpenAI models (GPT-6 Astra, or Soul) review the result.
- Why Theo cares
- OpenAI models dig into details and don't stop until they think they've found everything. They flag more things that don't matter, but sometimes catch real problems Claude models miss entirely.
- How to try it
- After Opus finishes a change, have an OpenAI model like Astra review it too.
- Expect some noise in the findings and triage what matters.
generally speaking, I find OpenAI models to be better at review still. They just really dig into the details and won't let go of the thing until it's confident it's found everything. It'll report more stuff that doesn't matter, but it will occasionally find things that do matter
entirely. So, personally, when I have Opus code being reviewed, I don't want just Opus reviewing it. I would like for Astra or even like Soul to give it a
Advice
Pre-approve what you know the agent will need, such as copying your environment variables (stored keys and settings). Otherwise it may worry, stop, and ask you for permission.
things like its ability to copy my environment variables. If I didn't tell it it could do that, it might be concerned when it goes to do that and then stop and ask me. I don't want it to
Explicitly tell Opus 5.5 it may use sub-agents (helper agents working in parallel). Theo finds it less willing to use them unprompted, so for big audits or piles of reviews he just tells it to.
saying here is that I've noticed Opus 55 seems less willing to spin up sub agents unless you tell it that it can. So I find myself often telling at the end of prompts, "Use sub agents and workflows however you choose. I trust your judgment there." Even then it doesn't
Ask the model to list what it could not verify itself, such as things needing other hardware or subjective calls. If you ask, it will tell you.
or it's a more subjective thing, if you ask the model to let you know what things it can't verify, it will. And
Use screenshots for feedback. Theo grabs a screenshot with a tool like Shutter, draws an arrow on the problem, pastes it in, and says 'This sucks. Make it better.' He says it works better than you'd expect.
tool like Shutter where I take a screenshot, I draw an arrow, I point it, and then I just paste this into the chat prompt like, I don't know, here, paste. And then I'll say something like, "This sucks. Make it better." That works way better than you would think. Next, we
Hot takes
Never use max reasoning on Opus 5.5. Use high or xhigh; he can barely tell them apart.
His reasoning: Other levels only cap how much the model may think. Max removes its ability to think less, forcing heavy thinking even on easy steps. High occasionally misses small details that xhigh catches; xhigh is a bit slower.
too hard. Max isn't just making it so the model can think more, it is removing its ability to think less. And that's what scares me about max reasoning, and why I specifically think you shouldn't
They're both fine. I don't have any issue with either, I can barely notice the difference with either. X high feels like it takes a little longer, high sometimes misses like smaller details that X high doesn't. Not that big of a
Today's models can work autonomously and verify their own changes, but they don't yet know how to work with a human. You have to tell them what you need.
His reasoning: The training data for working alongside people doesn't exist yet, and this way of working is still new.
stuff done. They can even verify their changes. They don't know how to work with a human yet. And that's partially because the data doesn't exist. It's partially because this is still a new phenomena. But if you tell the model how
How Theo thought about this
On a design comparison site, Fable 5.1 looked clearly better than Opus 5.5, yet others called Opus the better designer. Rather than pick a side, he worked out the gap: Opus isn't great at 'make it pretty' but follows design direction well.
But, I saw other people saying Opus was a way better designer, and I was confused because when I took this quick look here, didn't seem to be the case. Turns out, the thing Opus 5.5 is good at isn't just making a nice design when you say, "Hey, make it pretty." It's much better at going the right direction with
After his benchmark showed max taking 20 times longer in the worst cases for almost nothing, he asked whether max could still help, for example by not falling for trick answers. He grants 'perhaps' but still says never use max.
tokens, it took 20 times longer in the worst cases, and I got jack out of it. Are there things where max could help? Like maybe it is falling for a simple trick answer, but if it has to reason more, it doesn't fall for the trap. Perhaps. I still don't think you should ever use max reasoning. All the
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.
