Just Using Claude No Longer Generates Alpha
Eighteen months ago, simply transforming an ask into a Claude prompt was an edge. Your coworker would be dazzled by the magic of the LLM being able to spit out something coherent and dense almost immediately, almost unbelievably so. Even for those steeped in software and tech capabilities, AI was a quantum leap that was hard to fathom. Your role, as the humble Claude prompter, felt closer to a wielder of arcane witchcraft rather than an elevated Product Owner.
It’s an intoxicating experience, brokering someone’s first experience with capable AI. It is such a world-changing experience to have yourself; the immediate reaction is to find some way to chase that high again. Being an evangelist missionary, a pusher of the gateway drug, replicating that first encounter over and over, at any opportunity - it was like Lucy convincing her siblings to step through the wardrobe into Narnia.
Early apostles donned orange Claude-colored glasses. Have a feature you want to write? Claude. Presentation? Claude. Slack message? Claude. GitHub comment? Claude.
For a while, it worked. We were all hoping to find the limits - what drudgery couldn’t this eliminate, what timeline couldn’t be compressed? It surprisingly went farther than anyone could’ve guessed. Tokenmaxxing, usage awards, all signals nothing could (or should!) escape having the AI wand waved over it. Claude for coding, co-work, and even communication. It was a strange feeling at times, swapping simple PR review comments with a coworker using Claude to generate the actual text; it was the same username and smiling profile picture, but their clunky communication was as if they had been replaced by one of the Pod People, no longer actually communicating with the person.
When Claude first swept the workplace, just the fact that there was magic being used was more than enough. There was a premium for accessing a new technology whose abilities weren’t fully grasped - or even known - by most.
Since then, AI companies have run pretty incredible GTM strategies to multiplex AI tooling into form factors that fit any use case: video, desktop apps, voice, mobile, in-app, etc. The activation energy to use these products is close to zero. Access is no longer concentrated in the hands of early adopters; use of AI is no longer a distinguishing characteristic, but rather an expected competency. With the rapid adoption the expectations have risen. The professional behavior expected around AI usage has advanced pretty significantly. This both means an increase in the sophistication of usage - not only generating text but images, slide decks, in-line product features - but also eliminating the common anti-patterns that amateur users typically fall into.
Claude is table stakes today. Its simple magic alone no longer dazzles our audience. Everyone is aware, everyone has access. The alpha today is judgment on how, where, and when to use it; it is now magic that needs to be wielded appropriately.
We’re doubling down where it works well—places where outputs can be evaluated and verified without surrendering responsibility—as the capabilities still feel unbelievable even today. But we should hold ourselves and others accountable for the judgment behind the work. This matters especially in the more qualitative parts of work: understanding context, weighing tradeoffs, and communicating with other humans. Slop is easily recognizable. The days of blithely generating a massive AI-generated tome and throwing it over the wall for someone else to review are over. Simply inputting a plain prompt and passing along the output without refinement, verification, or even just looking at it for a moment is no longer acceptable. We know that the tooling can do that, because we’ve all been guilty of doing it ourselves; these days, after having been on the receiving end, we expect more.
The anti-patterns are probably pretty familiar: awkward Claude phrasing, verbosity with no real purpose, a work deliverable that is right in the letter of the law but clearly missed the spirit. Seeing these in the wild today unfortunately reflects negatively even on the most well-intentioned author. It’s sometimes hard to put a finger on why those anti-patterns feel so distasteful. I really, really like Sophie Alpert’s writing on this. It’s short, but attacks what I think the root of the problem is: writing is thinking, authorship is ownership, and precision is respect for others.
When an author delegates the expression of their ideas purely to AI, without editing or refinement, the audience must assume they are making a statement about their work. It may be that they don’t care about the intellectual ownership of their ideas - do they stand by every single word, phrase, idea in the output? It may be that they don’t really understand the task at hand and need to rely on AI as a crutch - and if they don’t understand the task at hand, can we trust what they’ve done? It may be that they are transferring the cognitive load - the hard task of editorial judgment - to the audience rather than owning it themselves. These are all assumptions, and none of them may be true; unfortunately, how the work is perceived ultimately matters more than the author’s actual intentions or competence. The audience only sees the output, not the author’s actual effort, and may assume the worst about the author’s judgment. Whether actually representative or just a case of misattribution, yielding editorial control to the AI tooling is a negative signal.
Done often enough, this cheapens the other good work that the author may be doing elsewhere. AI slop and well-intentioned but unedited AI output are indistinguishable at a glance. If it walks like AI slop and quacks like AI slop, it’s probably AI slop. And if it’s AI slop, the implication is that the author didn’t make the effort to think, that they didn’t really care what it actually said. Even with the most well-intentioned input, pure AI output erodes trust if it’s shared with others for them to seriously consume. It’s the same with any behavior that undermines trust within a team; the effect may not be felt immediately, but it will slowly sap confidence in a person’s ability to contribute meaningfully.
The judgment aspect of using AI is not one of mutual exclusivity - coding yes, writing no - but rather how to layer on AI use. It’d be silly (if not fully detrimental) to avoid use of AI during the thinking or communication process. When writing a blog post such as this, AI is still useful for research tasks or workshopping a clunky section or even getting some early feedback on an idea. AI should still be used when and where it can. The judgment lies in not completely yielding to its outputs or approaches, and in translating its output to fit the appropriate destination. AI is applied to reinforce, not replace, the hard parts of communication and critical thinking. The work that comes along with communicating an idea is part of fully understanding it - having to pick up the idea and look at it from the perspective of the target audience - and without that a crucial step is missed.
Not too long ago sharing AI output directly was bleeding edge. Today, that feels antiquated. It’s a sign that priors haven’t been updated, that the world has continued to move on, that Claude-speak might be replacing real thought. It’s a surprising turn of events that the very behavior that until recently signaled sophistication now risks signaling the opposite. I suspect, though, that the hard parts of thinking, context setting, and communication will continue to provide an edge as they traditionally have, even as the rest of how we work fundamentally shifts. Previously, just wielding the magic was enough; now, it is how to wield the magic that sets one apart.