current workflow to code with LLMs
my current workflow to code with LLMs
[[update: since summer 2026, agents needs way less hand-holding than described below]]
- provide a 200-lines context file to my project in
AGENTS.md - project objectives
- where to find relevant stuff (list of significant files)
-
preferences ("don't rebuild the project, as i'm already doing it")
-
have a clear mental model of what feature i want to build.
- i pretty much know how i would code it by hand
-
i know that the size / scope / complexity is manageable for the LLM
-
start dictating the new feature I want to work on. Dictating allows me to give more information and thus provide more context to the LLM. I've noticed that the LLMs are super good at sorting out my brain dump, and I get better results than typing it myself.
-
terminate my initial message with: "Before you do that, ask me clarifying questions. Number your questions (1, 2, 3) and label answer options (a, b, c, ...). Keywords are fine. " --> same as "Plan mode"
-
rinse and repeat until questions are trivial
-
type "go", let the LLM cook for a few minutes
-
review significant files (e.g. backend), skim medium important files (ts, html), skip a few (css)
-
run some tests
not using: - on stuff that i don't know about (LLM will run me in circles) - biggest models (latest Opus), unless nasty bug - Claude code: i like the concept, but prefer to be in an IDE where i can follow the process more - parallel agents (too much brute force to me) - vibe coding (:= not look at the code), unless for tiny 1-shot projects