Hun Labs

How I work

Since February 2026 I have handed almost all development to AI coding agents. So my work is less about writing code and more about making what the agents produce trustworthy. Across twenty-six repositories, six practices have stuck. Every example below comes from actual work records.

1Rules in documents, prohibitions in machines

Every repository has a rules file the agent reads before it starts. But irreversible actions don't stop at a line in a rules document: commands are checked mechanically before they run, and blocked.

Example In my first app, an agent deleted my rate-limiting logic and the core function of a filter while fixing something else. That led to a rules file and a change-log protocol (what changed, how to verify it, where the record is), and in the hardware I am building now it carried over into safeguards for places on the board, like its permanent-write areas, that can never be undone once written.

2A verification level on every claim

Claims that involve real hardware say how far they have been checked. What is only documented, what has been built, what the device responded to, what I confirmed with my own eyes, and what has been matched for accuracy are written in different words, and no level is skipped.

Example I raised the problem of verification claims that kept inflating; Claude Code proposed a format, and I chose it and spread it to the other repositories. The example code said the ring had thirteen LEDs; when I opened it up and counted, there were twelve. Colors that looked good on my monitor looked weak on the real ring, so I chose them again on the real thing.

3Predictions first

I write down predictions before seeing results, because once you have seen the result, any result can be explained. Wrong judgments are not erased; corrections are appended.

Example When the device's screen stuttered, I didn't fix it by guesswork: I wrote predictions first, then timed things inside the board to tell the causes apart.

4Audits through different eyes

An audit agent with no write access and AIs from other companies review the work separately. It carries over a standard I learned writing the paper: others must be able to check it again.

Example Each time I close out a piece of work, four different models read the raw conversation records separately and look for what the records missed or got wrong. I don't take any single AI's verdict as is; the final call I make myself, from logs and tests.

5Records outside the conversation

State does not live inside the conversation. Decisions and their grounds, a chronological log, and the point to pick up from are kept in files, and when the session changes, work restarts from there. Observed facts and interpretation are written separately.

Example This principle is also my research topic (the Cognitive Kernel Model, CKM). Because of it I could take out records a year old and use them again, and the story on this site was written from those records.

6Hunting for holes over adding safeguards

Having safeguards and having safeguards that work are different things. So I spend less time adding safeguards and more time keeping the cycle of finding holes and fixing them running.

Example A guard that blocks dangerous commands was checking only terminal commands, and I found out late that commands run through a plugin tool bypassed that check, so I turned the plugin off. Another time, a verification level that wasn't in the definitions table was used in the records for more than a day, until the four-model closing audit caught it and I added the definition.

How these practices came about is in Story.