Last week a teammate pushed back on a code review by pasting a Claude screenshot into the PR thread. The model had confidently explained why our current approach was wrong. The current approach was something I designed two years ago to satisfy a constraint that doesn’t appear anywhere in the diff Claude was looking at.

That’s been happening more often. A Slack reply that’s mostly a screenshot. A design doc with a “Claude says…” paragraph dropped in the middle. Someone in a meeting reading aloud from their phone. The framing is always a little apologetic and a little proud. Look what I found.

I use Claude every day. What I’m asking is narrow. Stop outsourcing your half of the conversation to it and expecting me to argue with the output.

#The cost is on me

When you paste a Claude screenshot at me, the time we spend stops being symmetric. You spent thirty seconds copying it. I spend twenty minutes reading the paragraph, finding the part that’s wrong, tracing it back to actual code, and writing a reply that explains why the model’s confident answer ignores something that matters.

Multiply that by ten engineers and a quarter goes by where I’m doing book reports on someone else’s prompt.

The model is a polished collaborator. It writes in complete paragraphs and it cites things, and the stuff it gets wrong is wrong in the same register as the stuff it gets right. Refuting a confident paragraph takes longer than writing one, and the model will write you ten more on demand.

#The model has never read our ERDs

Every team I’ve worked on has decisions that look weird from the outside, and properties of the code that aren’t visible in any single diff. The build pipeline does this odd thing because of a vendor contract that ended in 2023 but left scar tissue. The analytics SDK is wrapped in two layers of indirection because the privacy review in 2022 said it had to be. A base type’s protocol conformances are kept narrow on purpose because adding one at that layer cascades to every caller.

The model doesn’t know any of that. It’s read the public internet and your current diff. It’s going to suggest the textbook answer, and the textbook answer is exactly what we considered and rejected three years ago for reasons nobody is going to re-litigate in a PR comment. Solving the diff in front of you tells you nothing about how the rest of the system uses that code.

When you paste that answer at me as if it settles the question, you’re asking me to re-explain the last three years to a model that can’t remember the answer between turns.

#What I actually want from you

Tell me what you think. Even if you’re not sure. Especially if you’re not sure.

“I don’t get why we do it this way” is a great message. I can answer that one. “Here are two options I’m weighing, here’s what I’d lean toward” is even better. “I think this is wrong and here’s why” is the best of all, because now we’re having the conversation I can actually contribute to.

Your half-formed take tells me something a Claude paragraph can’t. It tells me what you think. If you’re wrong, I can correct the thing you’re wrong about. If you’re right and I disagree, you’ve taught me something. A polished Claude excerpt teaches me nothing about you, and I can’t push back on a model that isn’t in the room.

#Where Claude fits

Use Claude. I do. Use it to draft your message, check whether your reasoning has a hole, or summarize the thread you don’t want to read. Paste its output into your own scratchpad and pull out the bits that survive contact with what you actually believe.

Just don’t outsource your half of the conversation to it. The thing I want from you when we disagree is you.

Next time you’re about to paste a screenshot, try writing one sentence in your own voice first. Even if Claude said it better, your sentence is the one I can answer.¹