Sergey Kuhne stores AI coding tasks in git instead of open tabs: Reveneau

Why it mattersRunning more than one coding agent at once loses track of what each one is meant to do. Keeping tasks as files in git, with one open at a time, is a small pattern worth copying.

If you leave a coding agent running and open another tab to start a second task, you know this problem. By the fifth or sixth tab, you have forgotten what each agent was meant to do. Sergey Kuhne calls this "AI brain fry" in a blog post that reached 143 points on Hacker News. He is a solo founder working on Singing Carrots, an AI vocal coaching product.

His fix stores every work item as a markdown file in a git repository, next to the code. One agent cycles through the queue, and he only looks at one item at a time.

How the queue works

Each ticket is a markdown file with a status. Kuhne says the agent handles the slow parts. In his words, the agent does four things on each cycle. It researches new tickets. It builds the ones a human has tagged as safe to build. It runs the tests and a browser check. It verifies what is already deployed.

Decisions that need a person go into one inbox, in priority order. He opens the top ticket, decides, closes it, and moves on. Kuhne keeps the choices that need a person. What to build next, and whether to release to production.

The viewer, rebuilt from git

To see what the agent did, Kuhne opens a Claude artifact. In his words, the viewer is "one page, rebuilt from git on every cycle". The page carries the whole board plus one review page per ticket. Instead of scrolling back through a terminal, he opens one link. He reads what was built, and why.

The git repository is the real memory. The Claude artifact is a reading surface. Nothing lives in the agent's chat history. Chat history is lost the next time the session restarts.

What Kuhne is careful to say

Kuhne wrote his post as a personal note, not a research result. He says he was "hopping between 20 threads, trying to keep all those contexts in your brain RAM". This pattern is what helped him. He gives no numbers on how much time he saved. His closing line calls it "a huge step forward in my productivity with AI tools."

This is one person's workflow, posted to his own blog. The pattern is small. It uses only git, markdown files, a background agent and a Claude artifact as a reading surface.

For a founder or engineer running more than one AI agent at a time, there is one pattern worth trying. Every task is a file in git, with a status. Every decision closes a ticket. The chat tabs stay closed between decisions.

Source

How I'm fighting the AI Brain Fry, Sergey Kuhne. Discussion: Hacker News.

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.