2026-09-05 0003 AI Oriented Workflow

Start this semester, I joined a startup company to find a new direction for my life. And I was forced to setup an AI coding workflow on my local device. Be honest, this workflow is much more reliable than my imagnation:

Agent: OpenCode

Open Source Coding Agent, easy to use with TUI and can be directly installed by package manager in mainstream linux distros. (Deb, Fedora, Arch, Nix, ...)

Core Skill: Mattpocok Skills

This is the mainstream of the whole workflow, which puts engineers and LLMs in the right positions. It is not perfect, but it does help you to understand what LLM is actually doing.

LLM Model: deepseek-v4-flash / deepseek-v4-pro

This part is actually not that important as my friends thinking. With the previous skill, your LLM is expected to write code piece by piece. Just find the cheapest model which is able to handle such kind of work.

Version Control & Ticket Assign: Git + Linear

Git is a good management tool for codebase. But in AI workflow, share and arrange tickets in the whole team is also important, that is the reason for Linear. Compare to git, Linear's structure is more flatten and atomic, which is pretty friendly for context of LLM.

Currently this workflow works efficiently with only a little bit intervene from my decisions. But it also worries me, cause it runs so stable that my colleagues start to show unconsiousness about parts of the project. Maybe we can do better on it.