note

Coding agents and me

Written 27 days ago

Barely scratching the surface?

Automate everything!

A second coming of developers are happening. Everyone can now automate parts of their digital life.

But how should development teams work when everything can be 10x faster?

With new speed, one should slow down

I think we should shift our attention towards the earlier parts of creating software. Aligning earlier with product, designers and other developers, lets one utilize the new speed of coding without stopping to re-aligning (most of the time).

Early aligning for agents = aligning for devs

At a client we discussed how one can trust an agent more at correctly writing code. "Lots of our context lives in our heads. How do we make sure the agent does not break it?" "Lets write an agent.md file with these instructions"

This is something that developers ourselves even need. A team working document, a document with the expecations that the developers have for eachother when doing changes to the codebase. This includes formatting, style, interfaces, documentation, testing, PR review rules, and other expecations the team will have for eachother. This should be explicit, describing in straight details what is expected, maybe even providing examples. This can greatly help developers reduce friction thoughout writing and aligning around code.

My use of LLMs and coding agents

Github Copilot (auto-completion)

I picked up Github Copilot barely while they had some sort of free testing. Auto-completion by just tabbing through simple code blocks. Used to quickly create boilerplate or quickly writing up repeating patterns.

CSS block with all the 8(?) the client wanted all combinations of as a possibility in their coloring.

Nothing in between...?

Claude Code

I picked up Claude Code at the start of 2026. Coming into a project where most people would use AI for most things. Easy access and pretty liberal use for everything between making power points and write emails, to quick code review checks on PR's, test POC and early versions of feature.

Pi (pi.dev)

Something I came upon just now. Seems to follow the simple but highly extensible approach that neovim is also doing, which I like. I like this approach, as I can hone and whittle at the tool to fit exactly how I like (I should write about this)