Posts
- Ready to review
AI coding gets interesting before the coding starts
Recently I built a plugin that can take a task number, fetch the ticket, check the repository, and plan the work before writing a single line of code. I built it for Claude Code because I was tired of the same routine. Open the ticket. Read it twice. Search the codebase for the files it probably touches. Work out what's already there. Only then start. What surprised me is that the most useful part isn't the code it writes. It's the plan. When the AI has read the ticket and the repo first, the conversation changes. It stops guessing. It asks better questions. It points at the files that matter instead of inventing new ones. I think a lot of us use AI coding tools backwards. We ask them to write first and understand later. The interesting part of AI coding happens before the coding starts. Do you let AI plan before it writes, or does it go straight to code for you?
- New angle found
Agents won't remove the boring work. They'll expose it.
Everyone is reading the new agent announcements as “AI will do the boring parts of engineering now.” I think it's the other way around. Agents won't remove the boring work. They'll expose it. An agent that opens tickets, runs tests and updates statuses needs all of those things to be clear. Vague tickets, flaky tests, deploy steps that live in one person's head — humans quietly work around these every day. An agent just gets stuck. I've seen a small version of this with my own Claude Code plugin. The places where it struggled were exactly the places where our process was fuzzy. So before asking “what can agents do for us?”, it's worth asking “what would an agent trip over in our workflow?” That list is probably your real backlog.
- Draft improved
Stop asking AI to write code first
I think developers are using AI completely wrong. We open a chat and ask it to write code. A component, a hook, a migration. Then we spend the next twenty minutes fixing code that doesn't fit the project. The real leverage, at least for me, is asking it to read first. Before I let it write anything now, I ask three things: — Which files would this change touch? — What patterns does this codebase already use for this? — What could break? The answers are usually more useful than the code. AI is a decent writer of code. It's a much better reader of it.
- In progress
What async standups quietly took from our juniors
Three months ago we replaced our daily standup with a written async update. Fewer meetings. Better notes. Everyone got their mornings back. I'd still recommend it to most teams. But something changed that I didn't expect. Our junior developers started asking fewer questions.