A lot of posts about coding agents read like advertising. Not because anyone was paid, but because they follow the shape of advertising: here is a tool, here is what it did, here is how impressed I was.
The reader's defence goes up immediately, and the post does nothing for you.
Why it happens
The tool is the novel thing, so it becomes the subject. That is the mistake.
Nobody is short of information about what these tools can do. Everyone has seen the demo. What they have not seen is what happened when you used one on a real codebase with real constraints and a deadline.
The subject should be the work. The tool is context.
The rewrite
Before:
Been using a coding agent for a week and it is genuinely impressive. Built a whole feature in an afternoon. The productivity gains are real.
After:
Agent built the export feature in about twenty minutes. I then spent two hours on the part it got wrong, which was not code. It assumed exports were global. They are per-account. That constraint lives in one four year old ticket and three people's heads.
The slow part of AI assisted development is discovering which of your rules were never written down.
The second one has a finding. It is also less flattering to the tool, which is exactly why it reads as honest.
Three angles that hold up
Where the model was wrong in an interesting way. Not typos. Wrong assumptions that revealed something about your system.
What you had to explain that you thought was obvious. This is a documentation audit disguised as a workflow observation.
Where review broke down. Reviewing generated code is different from reviewing your own. Say how.
Three that do not
Speed claims without detail. "10x faster" means nothing and reads as sales copy.
Tool comparisons you did casually. Two afternoons is not a comparison, and people will argue with it rather than read it.
Anthropomorphising. "It got confused", "it decided". Cute once, tiresome quickly, and it obscures what actually happened.
The honesty test
Would you write this post if the tool's maker were never going to see it?
If the answer is no, the post is performing for someone. Write the version you would write in a private channel to a colleague who asked how it went.
That version is more specific, more critical, more useful, and it is the one people will actually read.
One more thing
If you use these tools, say so, on the posts where it matters. Not as a disclaimer on everything, but where the reader would want to know.
The alternative is being found out later, and the cost of that is higher than the cost of the sentence.