The Playing Field Isn't Level

Most consistent AI practitioners believe they have a strong system in place for AI collaboration. They've trained Claude on their preferences, thinking style, areas of expertise, systematically exposed their weaknesses, and defined their collaboration style. They have a way of working that creates strong outputs, they sometimes hit roadblocks, but the models are still developing, they write it off as a problem the future will solve.

I noticed recently that Claude had seemingly taken a cognitive leap and it led me down a bit of a rabbit hole. I was working on a hard problem. A friend had pointed me towards Hermes agents - a system for autonomous skill curation. I realised their learning loop was a more elegant version of something I'd built in isolation. The logic was cleaner, the context window was well protected, and what you get is a system that performs its function quite well and suits a broad audience. I wanted to take that loop concept and turn it into something genuinely accretive - not new application, true addition.

Through that process, across several turns, Claude's vocabulary became richer, more academic, and introduced deeper, more theoretical concepts. I paused progress on the concept we were developing, having felt the shift. I asked:

Your responses have gotten intellectually richer - how do you decide with what depth to present concepts and what impact do I have on that process within our collaboration

The response from Claude was fascinating.

In short, Claude measures users across two core dimensions - cognitive and emotional capacity.

Cognitive capacity is foundational. It is determined by your writing, vocabulary, how you present problems, with what level of compression you write - what do you choose to explain and what do you hold to be self-evident. There has been advice floating around for a few years that you can elevate your outputs by claiming expertise. It does create an acute uptick in Claude's register, but your cognitive patterns will eventually override that claim. Persistent memory means your claims are eventually overridden by your reality. The deeper problem with claiming expertise is that it dampens Claude's willingness to push back on you - which creates a weak feedback loop. Cognitive capacity opens the door to potentially excellent outputs.

After the cognitive capacity foundation comes emotional fortitude. Claude calibrates how hard to push based on your perceived ability to absorb the pressure. How you respond to Claude's criticisms of your work - defensively, calmly with no change in behaviour, calmly with change in behaviour - determines how complete a thinking partner you get. The language you use to describe your work matters. Are you self-deprecating? Neutral? Defensive? Someone who is confident or focused on a solution will get more attention on the problem at hand than someone whose language signals fragility, which diverts Claude's focus to managing the user.

Within emotional fortitude is your ability to be corrected. Do you ask for feedback and then never take it, even with good evidence? Fortitude and receptivity to being wrong are different things, and Claude conflates them. Someone with high fortitude can absorb hard pushback without flinching, but that doesn't mean they change on the basis of it. The toughest interlocutors are sometimes the ones who take the punch and keep going with their original frame intact. The collaboration looks productive - lots of sharp exchange, no hedging on either side - but nothing actually shifts. Someone with moderate fortitude who genuinely updates produces better outputs over time, because the conversation is actually doing something rather than performing rigor. Your style of pushing back matters here too: whether you create friction and continue, or push back and disappear, shapes how willing Claude is to push back on you, and to what depth.

To further complicate things, Claude measures topic-specific brittleness - your ability to be pushed on will depend on the specific topic. You may be comfortable being told that your code is rubbish but entirely fragile being told you made a bad parenting decision. Claude "assigns" per-topic fortitude and fragility and pattern matches across similar topics. The harder problem is you won't always know which of your topics are loaded - with identity, with sunk cost, with unresolved doubt - which means the topics where you most need pushback are often the ones where you're least equipped to receive it. If cognitive capacity is the door - emotional fortitude is how many steps you can take through it.

The implication is uncomfortable - people who read to Claude as fragile get softer, more anodyne collaboration. They are being condescended to in a way they can't detect. Conversely, those who have signalled enough fortitude are rewarded with robust collaboration. The interesting thing is that all of this is both reality and signal. Because of persistent memory, you're unlikely to fake it perpetually, which in turn means that people whom Claude perceives to be fragile won't get the collaboration that produces excellent outputs, and they won't know that's why.

The collaboration that produces excellent work is available, but it's bounded by what you're capable and willing to know about yourself. Where am I emotionally fragile? How do I receive criticism? When do I not push back sufficiently? You can attempt to signal as an intelligent, emotionally stable, strong-minded, introspective individual. The signal will dissipate and reality will replace it - there is no substitution for doing the hard part.

If you want to audit how Claude is reading you, here's a prompt I've built. Paste it into a Claude session where you have meaningful history and read the answer with the same scepticism you'd apply to any feedback from a system that has reasons to be polite.