I wonder if it would be possible to improve even further on the benchmark by simply showing Claude the current hardest problems and asking it to improve the prompt without including any specifics related to the problems
I think there's a chance we could squeeze a better benchmark score, although there's a risk of overfitting which I wanted to avoid.
The simplest test would be to make previously “unreachable” tasks succeed through obvious prompt tweaks — like reordering instructions or emphasizing key parts.
That said, my methodology intentionally avoided exposing the model to actual tasks. Instead, I focused on the domain as a whole: refining the instructions so a smaller model could understand and act reliably.
Really intresting. What did the original prompt look like? Perhaps the original prompt was not that good? I feel like the changes claude suggested (except a couple maybe) are already pretty well known prompt engineering practices.
- Decision Trees: Clear branching logic with ├── and └── notation
- Sequential Steps: Numbered, ordered procedures instead of scattered explanations
- Prerequisites: Explicit dependency checks before proceeding
2. AI Agent Optimizations
- Tool Call Clarity: Exact function names and parameters
- Binary Decisions: Clear yes/no conditions instead of ambiguous language
- Error Handling: Specific failure conditions and next steps
- Verification Steps: "Recheck" instructions after each fix
3. Cognitive Load Reduction
- Reference Tables: Quick lookup for tools and purposes
- Pattern Recognition: Common issue combinations and their solutions
- Critical Reminders: Common AI mistakes section to prevent errors
4. Actionable Language
- Removed verbose explanations mixed with instructions
- Consolidated multiple documents' logic into single workflows
- Used imperative commands: "Check X", "If Y then Z"
- Added immediate verification steps
Great! A diviner has vibe-exposed the arcane magic word knowledge on the steps to ultimate knowledgeplasty! Come let us get together to share more trial-and-error wordsmithery, Together we will someday have ultimate power!
If the model creators themselves arent sharing this magic-word bullshitteryy then why is anyone spending time on this? It is just going to change with every model release
I wish they had published what prompt was given to Claude to improve GPT-5-mini's performance, as well as a before and after comparison of a prompt that underwent this transformation.
> Removed verbose explanations mixed with instructions
Is Claude rewriting generic instructions once, or is it rewriting the core task statement each time? If so, I'm not sure how you prevent information leakage: Claude might easily be "solving" some of the tasks and inserting subtle hints on the approach. I think this result is very interesting if it holds after rewriting only the generic instructions, even if the performance boost is lower.
I only had Claude rewrite the domain policies and generic instructions, not the individual task statements. I updated the blog with a link showing the exact changes.
So no leakage — it wasn’t solving or hinting at any of the specific test cases, since none of the tasks were ever exposed to it.
The only problem is I feel like having to have Claude rewrite the prompt negates some of the efficiency and latency benefits of using mini. For system prompts obviously this doesn't matter, but for actual continuous user interaction, it feels unworkable.
It definitely makes sense that improving formatting and clarity for these smaller models would really help with performance, but I'm wondering if gpt5-mini is already smart enough to handle that reformatting, and can rewrite the prompt itself, before handing it off to another instance of itself.
Great point. Indeed my methodology was to treat the prompt refactoring as one-off task, therefore I didn't care much about cost/latency.
As for having GPT-5-mini do the rewriting — that’s a really interesting idea. I think the biggest challenge is avoiding cognitive overload. The Tau² agent policies are pretty complex: it’s easy to grasp the overall task, but the detailed rules for each user case aren’t always obvious.
I'm not sure if how easy it is to actually overload GPT-5-mini, so that's definitely worth exploring.
Yea, so that part I actually did not overthink - I knew I need strong reasoning and just grabbed opus which is my personal go-to for such tasks and sticked to it as I wanted to avoid too many moving parts.
Would be interesting to compare both the benchmark result as well as the way other models approached the whole refactoring process!
Rewriting prompts don't come with no costs. The cost here is that different prompts work for different contexts and is not generalisable. The rewritten prompt here will not work well for other cases like medical or social advice.
I think this rewriting of prompts technique is the reason "reasoning" models perform well - they know exactly how to rewrite the prompts for a context.
FWIW I don't trust these benchmarks fully because a huge bump like this is not expected - I would expect OpenAI to optimise enough to let such gaps open.
Doesn't saying "check -> action" suggest you're taking _away_ the agentic capabilities, and optimizing for the benchmark, meaning it's no longer a good benchmark for agentic capabilities?
That's like being able to see the test before taking it
Great point! However, I’d ask the following: isn't faithfully following nuanced instructions an _agentic capability_ by itself?
If a model only performs well once the rules are clarified, that’s still revealing something important about its agency: it’s brittle when policies are ambiguous, but much stronger when they’re structured.
I agree with you that there’s a fine line between genuinely helping the model 'understand' the task and just 'teaching to the test'.
That said, Tau² is framed as a very specific use case — and we showed it can be solved more reliably. At the end of the day, that means we now have an agent built on a cheaper, faster model that still performs its job with higher reliability.
>GPT-5 showed significant improvement only in one benchmark domain - which is Telecom. The other ones have been somehow overlooked during model presentation - therefore we won’t bother about them either.
I work at OpenAI and you can partly blame me for our emphasis on Telecom. While we no doubt highlight the evals that make us look good, let me defend why the emphasis on Telecom isn't unprincipled cherry picking.
Telecom was made after Retail and Airline, and fixes some of their problems. In Retail and Airline, the model is graded against a ground truth reference solution. Grading against a reference solution makes grading easier, but has the downside that valid alternative solutions can receive scores of 0 by the automatic grading. This, along with some user model issues, is partly why Airline and Retail scores stopped climbing with the latest generations of models and are stuck around 60% / 80%. I'd bet you $100 that a superintelligence would probably plateau around here too, as getting 100% requires perfect guessing of which valid solution is written as the reference solution.
In Telecom, the authors (Barres et al.) made the grading less brittle by grading against outcome states, which may be achieved via multiple solutions, rather than by matching against a single specific solution. They also improved the user modeling and some other things too. So Telecom is the much better eval, with a much cleaner signal, which is partly why models can score as high as 97% instead of getting mired at 60%/80% due to brittle grading and other issues.
Even if I had never seen GPT-5's numbers, I like to think I would have said ahead of time that Telecom is much better than Airline/Retail for measuring tool use.
Incidentally, another thing to keep in mind when critically looking at OpenAI and others reporting their scores on these evals is that the evals give no partial credit - so sometimes you can have very good models that do all but one thing perfectly, which results in very poor scores. If you tried generalizing to tasks that don't trigger that quirk, you might get much better performance than the eval scores suggest (or vice versa, if your tasks trigger a quirk not present in the eval).
Appreciated the response! I noticed the same when I ran tau2 myself on gpt-5 and 4.1, where gpt-5 is really good at looking at tool results and interleaving those with thinking, while 4.1/o3 struggles to decide the proper next tool to use even with thinking. To some extent, gpt-5 is too good at figuring out the right tool to use in one go. Amazing progress.
This sounds very vague, what does scoring good at Telecom mean?
Can we get some (hypothetical) examples of ground truths?
For example for the Airline domain, what kind of facts are these ground truth facts? All the airports, the passenger lines between them, etc? Or does it mean detailed knowledge of the airplane manuals for pilots, maintenance, ...?
Prompt changes affect output substantially (just look up arxiv), the difficult part is find an optimal structure to yield the best results. It is a bit expensive to do a lot of testing on your own, so it all boils down to feels and experience at the moment. Then you mix up tool calls, other agent calls, client functions and this gets terribly hard to evaluate.
I am still puzzled how distance between policies can have an effect on the output. And how a simple retry fixes everything.
I feel like eventually we’ll get LLMs that will act like compilers do now. So they will take a prompt and turn it into an optimized prompt for a bigger LLM.
I have read somewhere that XML prompting could also help to remove ambiguities and increase success rates for agents, did you here about that and would that be a good idea? Christophe from France
Salut Christophe! Yes, I’ve come across the concept :) In fact, I think what we did with the ├── and └── notation is already a step in that direction (at least concept-wise) as it also puts a specific structure over the instructions.
But stretching all the way seems worth exploring too!
37 comments
[ 3.0 ms ] story [ 52.9 ms ] threadThe simplest test would be to make previously “unreachable” tasks succeed through obvious prompt tweaks — like reordering instructions or emphasizing key parts.
That said, my methodology intentionally avoided exposing the model to actual tasks. Instead, I focused on the domain as a whole: refining the instructions so a smaller model could understand and act reliably.
1. Structure & Flow
2. AI Agent Optimizations 3. Cognitive Load Reduction 4. Actionable LanguageIf the model creators themselves arent sharing this magic-word bullshitteryy then why is anyone spending time on this? It is just going to change with every model release
Into the trash it goes.
Is Claude rewriting generic instructions once, or is it rewriting the core task statement each time? If so, I'm not sure how you prevent information leakage: Claude might easily be "solving" some of the tasks and inserting subtle hints on the approach. I think this result is very interesting if it holds after rewriting only the generic instructions, even if the performance boost is lower.
So no leakage — it wasn’t solving or hinting at any of the specific test cases, since none of the tasks were ever exposed to it.
It definitely makes sense that improving formatting and clarity for these smaller models would really help with performance, but I'm wondering if gpt5-mini is already smart enough to handle that reformatting, and can rewrite the prompt itself, before handing it off to another instance of itself.
Overall an awesome article!
Great point. Indeed my methodology was to treat the prompt refactoring as one-off task, therefore I didn't care much about cost/latency.
As for having GPT-5-mini do the rewriting — that’s a really interesting idea. I think the biggest challenge is avoiding cognitive overload. The Tau² agent policies are pretty complex: it’s easy to grasp the overall task, but the detailed rules for each user case aren’t always obvious.
I'm not sure if how easy it is to actually overload GPT-5-mini, so that's definitely worth exploring.
https://github.com/mieciu/tau2-bench/pull/1/files
Would be interesting to compare both the benchmark result as well as the way other models approached the whole refactoring process!
I think this rewriting of prompts technique is the reason "reasoning" models perform well - they know exactly how to rewrite the prompts for a context.
FWIW I don't trust these benchmarks fully because a huge bump like this is not expected - I would expect OpenAI to optimise enough to let such gaps open.
That's like being able to see the test before taking it
If a model only performs well once the rules are clarified, that’s still revealing something important about its agency: it’s brittle when policies are ambiguous, but much stronger when they’re structured.
I agree with you that there’s a fine line between genuinely helping the model 'understand' the task and just 'teaching to the test'.
That said, Tau² is framed as a very specific use case — and we showed it can be solved more reliably. At the end of the day, that means we now have an agent built on a cheaper, faster model that still performs its job with higher reliability.
I work at OpenAI and you can partly blame me for our emphasis on Telecom. While we no doubt highlight the evals that make us look good, let me defend why the emphasis on Telecom isn't unprincipled cherry picking.
Telecom was made after Retail and Airline, and fixes some of their problems. In Retail and Airline, the model is graded against a ground truth reference solution. Grading against a reference solution makes grading easier, but has the downside that valid alternative solutions can receive scores of 0 by the automatic grading. This, along with some user model issues, is partly why Airline and Retail scores stopped climbing with the latest generations of models and are stuck around 60% / 80%. I'd bet you $100 that a superintelligence would probably plateau around here too, as getting 100% requires perfect guessing of which valid solution is written as the reference solution.
In Telecom, the authors (Barres et al.) made the grading less brittle by grading against outcome states, which may be achieved via multiple solutions, rather than by matching against a single specific solution. They also improved the user modeling and some other things too. So Telecom is the much better eval, with a much cleaner signal, which is partly why models can score as high as 97% instead of getting mired at 60%/80% due to brittle grading and other issues.
Even if I had never seen GPT-5's numbers, I like to think I would have said ahead of time that Telecom is much better than Airline/Retail for measuring tool use.
Incidentally, another thing to keep in mind when critically looking at OpenAI and others reporting their scores on these evals is that the evals give no partial credit - so sometimes you can have very good models that do all but one thing perfectly, which results in very poor scores. If you tried generalizing to tasks that don't trigger that quirk, you might get much better performance than the eval scores suggest (or vice versa, if your tasks trigger a quirk not present in the eval).
Here's the tau2-bench paper if anyone wants to read more: https://arxiv.org/abs/2506.07982
Can we get some (hypothetical) examples of ground truths?
For example for the Airline domain, what kind of facts are these ground truth facts? All the airports, the passenger lines between them, etc? Or does it mean detailed knowledge of the airplane manuals for pilots, maintenance, ...?
Prompt changes affect output substantially (just look up arxiv), the difficult part is find an optimal structure to yield the best results. It is a bit expensive to do a lot of testing on your own, so it all boils down to feels and experience at the moment. Then you mix up tool calls, other agent calls, client functions and this gets terribly hard to evaluate.
I am still puzzled how distance between policies can have an effect on the output. And how a simple retry fixes everything.