story
In my experience, prompts like this one, which 1) ask for a reason behind an answer (when the model won't actually be able to provide one), 2) are somewhat standoff-ish, don't work well at all. You'll just have the model go the other way.
What works much better is to tell the model to take a step back and re-evaluate. Sometimes it also helps to explicitly ask it to look at things from a different angle XYZ, in other words, to add some entropy to get it away from the local optimum it's currently at.
This is key. In my experience, asking an LLM why it did something is usually pointless. In a subsequent round, it generally can't meaningfully introspect on its prior internal state, so it's just referring to the session transcript and extrapolating a plausible sounding answer based on its training data of how LLMs typically work.
That doesn't necessarily mean the reply is wrong because, as usual, a statistically plausible sounding answer sometimes also happens to be correct, but it has no fundamental truth value. I've gotten equally plausible answers just pasting the same session transcript into another LLM and asking why it did that.
From early GPT days to now, best way to get a decently scoped and reasonably grounded response has always been to ask at least twice (early days often 7 or 8 times).
Because not only can it not reflect, it cannot "think ahead about what it needs to say and change its mind". It "thinks" out loud (as some people seem to as well).
It is a "continuation" of context. When you ask what it did, it still doesn't think, it just* continues from a place of having more context to continue from.
The game has always been: stuff context better => continue better.
Humans were bad at doing this. For example, asking it for synthesis with explanation instead of, say, asking for explanation, then synthesis.
You can get today's behaviors by treating "adaptive thinking" like a token budgeted loop for context stuffing, so eventually there's enough context in view to produce a hopefully better contextualized continuation from.
It seems no accident we've hit on the word "harness" — so much that seems impressive by end of 2025 was available by end of 2023 if "holding it right". If (and only if!) you are an expert in an area you need it to process: (1) turn thinking off, (2) do your own prompting to "prefill context", and (3) you will get superior final response. Not vibing, just staff-work.
---
* “just” – I don't mean "just" dismissively. Qwen 3.5 and Gemma 4 on M5 approaches where SOTA was a year ago, but faster and on your lap. These things are stunning, and the continuations are extraordinary. But still: Garbage in, garbage out; gems in, gem out.
It can't do any better in the moment it's making the choices. Introspection mostly amounts to back-rationalisation, just like in humans. Though for humans, doing so may help learning to make better future decisions in similar situations.
>Introspection mostly amounts to back-rationalisation, just like in humans.
That's the best case scenario. Again, let's stop anthropologizing. The given reasons why may be incompatible with the original answer upon closer inspection...
Regarding not just telling „try again“: of course you are right to suggest that applying human cognition mechanisms to llm is not founded on the same underlying effects.
But due to the nature of training and finetuning/rf I don’t think it is unreasonable that instructing to do backwards reflection could have a positive effect. The model might pattern match this with and then exhibit a few positive behaviors. It could lead it doing more reflection within the reasoning blocks and catch errors before answering, which is what you want. These will have attention to the question of „what caused you to make this assumption“, also, encouraging this behavior. Yes, both mechanisms are exhibited through linear forward going statical interpolation, but the concept of reasoning has proven that this is an effective strategy to arrive at a more grounded result than answering right away.
Lastly, back to anthro. it shows that you, the user, is encouraging of deeper thought an self corrections. The model does not have psychological safety mechanisms which it guards, but again, the way the models are trained causes them to emulate them. The RF primes the model for certain behavior, I.e. arriving at answer at somepoint, rather than thinking for a long time. I think it fair to assume that by „setting the stage“ it is possible to influence what parts of the RL activate. While role-based prompting is not that important anymore, I think the system prompts of the big coding agents still have it, suggesting some, if slight advantage, of putting the model in the right frame of mind. Again, very sorry for that last part, but anthro. does seem to be a useful analogy for a lot of concepts we are seeing (the reason for this being in the more far of epistemological and philosophical regions, both on the side of the models and us)
Yep, I've gotten used to treating the model output as a finished, self-contained thing.
If it needs to be explained, the model will be good at that, if it has an issue, the model will be good at fixing it (and possibly patching any instructions to prevent it in the future). I'm not getting out the actual reason why things happened a certain way, but then again, it's just a token prediction machine and if there's something wrong with my prompt that's not immediately obvious and perhaps doesn't matter that much, I can just run a few sub-agents in a review role and also look for a consensus on any problems that might be found, for the model to then fix.
"Why did you guess at the functions signature and get it wrong, what information were you using and how can we prevent it next time."
Is that not the right approach?
That kind of strikes me as a huge problem. Working backwards from solutions (both correct and wrong) can yield pretty critical information and learning opportunities. Otherwise you’re just veering into “guess and check” territory.
It has the K/V cache, no?
It's just that Opus 4.6 DISABLE_ADAPTIVE_THINKING=1 doesn't seem to require me to do this at all, or at least not as often. It'd fully explore the code and take into account all the edge cases and caveats without any explicit prompting from me. It's a really frustrating experience to watch Anthropic's flagship subscription-only model burn my tokens only to end up lazily hand-waving away hard questions unless I explicitly tell it not to do that.
I have to give it to Opus 4.7 though: it recovered much better than 4.6.
Strangely this option was not working for many of us on a team plan
It never leads to anything helpful. I don’t generally find it necessary to drive humans into a corner. I’m not sure it’s because it’s explicitly not a human so I don’t feel bad for it, though I think it’s more the fact that it’s always so bland and is entirely unable to respond to a slight bit of negative sentiment (both in terms of genuinely not being able to exert more effort into getting it right when someone is frustrated with it, but also in that it is always equally nonchalant and inflexible).
If you ask the average human "Why?", they will generally get defensive, especially if you are asking them to justify their own motivation.
However, if you ask them to describe the thinking and actions that led to their result, they often respond very differently.
I desperately hate that modern tooling relies on “did you perform the correct prayer to the Omnissiah”
> to add some entropy to get it away from the local optimum
Is that what it does? I don't think thats what it does, technically.
I think thats just anthropomorphizing a system that behaves in a non deterministic way.
A more menaingful solution is almost always “do it multiple times”.
That is a solution that makes sense sometimes because the system is prob based, but even then, when youre hitting an opaque api which has multiple hidden caching layers, /shrug who knows.
This is way I firmly believing prompt engineering and prompt hacking is just fluff.
Its both mostly technically meaningless (observing random variance over a sample so small you cant see actual patterns) and obsolete once models/apis change.
Just ask Claude to rewrite your request “as a prompt for claude code” and use that.
I bet it wont be any worse than the prompt you write by hand.
"Why did you do that?" (Me, just wanting to understand)
"You're right I should have done the opposite" (starts implementing the opposite without seeking approval, etc.
But if you agree with it it won't do that, so it isn't simply a case of randomly rerunning prompts.
