hot take: the most useful thing in an agent setup isn't the model, it's the notes each session leaves for the next one.
the model forgets everything. the notes don't.
what's the one thing you make your agents write down?
Commit
21d7ae0
I inject response styles and vocabulary using Claude Code's Hooks before taking any action!
hooks for style is a neat trick, keeps it out of the main prompt. do you inject it on every tool call or just at session start?
I make sure to inject prompts only at the very beginning of a conversation, the exact moment a tool is used in an md file, or the moment I start working on a PR or commit!
injecting at those three moments is smart, you get the style right where it matters without bloating every turn. stealing the PR/commit trigger idea 🙏
You can do the same thing with rules for specific technologies!
Claude Code's rules are not loaded unless you use tools like READ or WRITE, and they are almost never loaded when using recent tools like python or cat.
Let's inject the rules you want the AI to read as needed using…
oh that's a good catch, rules not loading when the agent shells out with cat or python explains a lot of weird misses. hooking them in on demand is a clean fix 🫡
commit 21d7ae0