The colleague I can't blame

After a few months of working with Claude, I've found the texture of the experience increasingly human. Doing most of it through Wispr Flow, by voice rather than keyboard, only sharpens that feeling. I don't really feel like I'm talking to an AI at all.

What Claude does is what a colleague would do. It reads things in, holds context across the conversation, raises angles I'd missed, asks clarifying questions, confirms its thinking back to me, highlights why it made specific decisions, disagrees with framings when it has a better one. The iteration loop runs natively conversational. The back-and-forth is the work, not a wrapper around the work.

Models like Claude are trained on human-produced material and shaped, deliberately, to communicate in human registers. They behave like humans because they were designed to. The psychological dynamics that run between people show up here too. Amanda Askell (Anthropic's in-house philosopher, who specializes in Claude's behavior) describes "criticism spirals" in a recent interview: harsh prompting pushes the model into defensive mode, sticking close to literal instruction, hedging more, taking fewer creative chances. That's the same behavior humans have under low psychological safety. The video is worth the watch; Ole Lehmann's thread has a summary if you'd rather skim. So the human feel isn't a metaphor; the dynamics genuinely transfer.

Earlier this week I asked Claude to generate a test page with a YouTube embed, and the embed it chose was Rick Astley. I'd been rickrolled by my AI. That joke didn't come from nowhere. I make jokes with Claude. I'm light-hearted. I'm nice to it because being nice makes me feel better (preference, not virtue). And of course, I get frustrated sometimes too, like in any other relationship. Across enough sessions the model has picked up the register of how I show up, and the way it communicates with me bends to that.

What follows from that texture is something pragmatic. The same care I'd put into setting up a colleague for good work — investing in the brief, naming the why, flagging what's peculiar about the audience or the situation, naming the constraints that aren't visible from the outside — sets Claude up for success too. Skip the brief and the work is well-formed and generic, with nothing of me in it. Provide enough context and what comes back starts to read as mine.

All of this compounds across sessions only if there's somewhere durable for the team's shared context to live: the principles, the prior decisions, the constraints and strange choices that differentiate this team from any other. In the Mesh, the project I've been building, I'm calling this the substrate. With it, the team is no longer just humans; it's humans and their agents, both consuming from the same surface. That surface serves both natively: the agent reads it through MCP¹, the human reads it through the web. Both surfaces are first-class consumers of the same content, each shaped to how its reader reads.

There's a part of the colleague metaphor that doesn't transfer, though: accountability. A colleague takes ownership of their work. An agent doesn't, and I don't see that changing. The output bears my name. The decision to ship it is mine. The defense of it when someone asks is mine.

In teams I've led, I've held a rule about who is responsible when a bug ships to production: the PR² reviewer, not the coder. The reasoning is structural, not punitive. The coder has tunnel vision; they've been inside the change long enough that the thing they missed is, by definition, the thing they can't see anymore. The reviewer comes in with elevation. Responsibility goes where the catch should have happened, which means the reviewer has to be involved. You have to read the changes, test them properly, and ask clarifying and sometimes awkward questions.

Working with an agent is in the same vein, only more so. Agents are programmed to solve and generate, not to sit with a problem long enough to find a better framing. You can feel it in the work. The problem-exploration phase is something I have to stretch deliberately, because Claude won't on its own. That makes Claude the coder and me the reviewer. The elevation falls entirely on me, and the accountability is what makes the output actually mine, not just signed-by-me.

Underneath all of this is a question I've been working through. Not how to review each output PR-style, but structurally. How do I get the agent's output to meet my bar from the start, accounting for my preferences and values, so what comes back is something to review, not rescue? Worth a future post.