profile

Not a vibe coder

Sub Agents

Published 11 days ago • 1 min read

Sub Agents

I’ve been playing around with Claude Code sub-agents a lot recently and only just cracked how to get them working reliably.
​
At first I thought they’d just work automatically. After all, the docs do say just prompt Claude with ALWAYS USE and it’ll magically select the right agent.
​
But that never worked for me.
​
Instead I found tagging the agent with @sub-agent-name forced Claude to select it. This feels a bit broken but I guess its messy tool selection and thats a tricky problem.
​
I discovered a better method.
​
Asking Claude to tag the right agent when it builds the plan. This fits really nicely in my current workflow:
​
→ research
→ plan (this step)
→ execute
→ refactor
​
It turns out to be a really nice way to get it to automatically add agent selection to the planning stage without me having to manually select the agent at different stages. Its not full-proof but it does it a good majority of the time.
​
I should probably spend more time tweaking the agent descriptions and Claude.md file to get it to eagerly select these agents but I haven’t invested the time there yet.
​
I’ll have a go and make a post when I test this.

113 Cherry St #92768, Seattle, WA 98104-2205
​Unsubscribe · Preferences​

Not a vibe coder

Master AI Coding Tools

I'm sharing the exact methods I use as an AI engineer to build production systems - techniques so effective they work even in niche languages like Elixir. Get the complete playbook delivered to your inbox 👇

Read more from Not a vibe coder

Hooks Hooks are another feedback mechanism we can use in Claude code.They run arbitrary commands after certain lifecycle events.There’s one hook I lean on heavily “Stop”.It’s called when Claude has finished generating code. Which maps really nicely to my normal workflow where I would run commands like compile, lint, prettier, etc. to check the code is in a good state before continuing.These commands are generally fast to run and when they fail the error is fed back to the LLM and continue to...

11 days ago • 1 min read

Background Agents My favourite thing about Github Agents is that they run in their own isolated containers and can scale infinitely. Until recently I’ve felt a bit stuck with Claude Code. On one hand its the best code generator. On the other, it doesn’t scale well beyond one instance and I have to wait for it to complete, which sometimes can be quite a while. Now I’ve figured out a way to work on multiple tickets at a time by running multiple instances of claude code. This is something they...

11 days ago • 1 min read