I kicked off a release notes pipeline this week that was built to run end to end with no one watching: pull tags and git logs across three repos, cross-reference eighteen Jira tickets, write the summary, publish it. Fully wired. Fully automated.
We stopped putting the same model behind every agent this week. The expensive one decides what to do now, and a cheaper one does it.
Our AI Steering Committee met this week and, among other things, decided we should stop using Claude to build internal-facing slide decks. If the audience is internal, make it a memo instead.
There’s a move I’d make to get an engineering team to take AI seriously, and it’s the opposite of what most orgs reach for. Don’t announce the tool. Show it, live, on the code the team shipped this week.
I had the name before I had a single line of code. Codegraph. I’d spent real time on it: what it would do, who it was for, what the first version needed to ship. Then I sat down to build and found out it already existed. A public repo, thousands of stars, actively maintained, doing most of what I’d just finished scoping.
The post you’re reading took multiple rounds with an LLM before I was willing to put my name on it. Not one prompt and a paste: pass after pass of me handing the draft back. Cut that sentence. That transition is carrying nothing. I don’t talk like that. The punctuation tics are a tell, take them out. The point’s buried in the middle, bring it up top.
Midway through generating release notes for a recent release, the agent hit its context limit. Not a crash, not an error: it just ran out of room to keep the whole task in its head at once.
There’s a discipline I’m trying to enforce on how we use AI, and the vendor makes it inconvenient on purpose: keep the context in a tool they don’t own.