Ai on @zhisme :: signal over noise

What Is A Harness, Demystified

Fri, 07 Aug 2026 00:00:00 +0000

One fine day a tiny laddie
came and asked his dad:
"Could you tell me, Daddy,
what is good and what is bad?"

V.V. Mayakovsky

"What Is Good And What Is Bad" (1925), trans. Dorian Rottenberg

From testing

The word isn’t new. Test harness¹ is a thirty-year-old term: the rig that runs the code under test, feeds it inputs, collects outputs, and says “pass / fail”. Swap “code under test” for “model” and you get exactly what everyone is talking about now.

LLM Strikes Back But Not The Way We Wanted

Fri, 29 May 2026 00:00:00 +0000

AI strikes back, but not the way we wanted.

Recently I had an interview with a recruiter — a very interesting and unexpected interview for me. It’s a funny story that can teach you something and make you vigilant.
The interview started in a very cliché way: asking me about my experience, why I left my previous company, what I’m looking for, and what my goals are for a new workplace.

Give Your LLM Brevity

Thu, 23 Apr 2026 00:00:00 +0000

Brevity is the Sister of Talent

A.P. Chekhov

Why?!

I don’t want to read invented machinery poems from an LLM about why it chose BigDecimal(0.5) instead of a literal 0.5 and writes 3 paragraphs defending a choice I didn’t ask about.
I don’t want to read invented nonsense that I have no time for.
If I don’t like something, I’ll ask to reimplement/refactor, or in rare cases request elaboration.
Then I will cover it with the needed tests until it satisfies me.
If you’re like me, you don’t want to spend time on that.

Practical Useful Commands for Claude

Tue, 17 Mar 2026 00:00:00 +0000

These are commands I use often and see a lot of value from. I advise you at least to consider using them in case you haven’t tried them yet.

Commands

claude --worktree or claude -w

In case you want multiple agents to run simultaneously, this option was designed for it. Each Claude agent gets its own version of the repo so they won’t interfere with each other. Where can you use it? Let’s say you’re migrating something to a new architecture and there are multiple services that need changes, but they’re located in a monorepo. With this option, each agent will be moving its own service or “folder” if you want, without getting conflicts with others.

Disable Co-Authored-By Claude in Git Commits

Fri, 27 Feb 2026 00:00:00 +0000

Solution

Open ~/.claude/settings.json with your favorite editor and add this to your json config. This is taken from Claude docs¹

{
 // ...
 "attribution": {
 "commit": "",
 "pr": ""
 }
}

Why

This actually makes no sense and I agree with Rob Pike’s take on that². It just adds non-readable boilerplate that does not add any value to the git commit. If it was committed by some guy I expect him to take responsibility of that commit, not some excuse like “that was written by Claude, sorry I didn’t check”. So it makes no sense at all, we are blind to that string anyway. Git commits should be atomic, simple and readable, not overwhelmed by nonsense. The trick above will help you with that.

OpenClaw Chameleon Hype

Wed, 11 Feb 2026 00:00:00 +0000

The Chameleon With Claws

clawdbot/moltbot/OpenClaw — this chameleon flew through our tight-knit tech world and generated a spectacular amount of hype. 184k GitHub stars in the last month alone. For context, Rails has 56k stars after nearly 20 years.

So what awaits this clawed chameleon — a rapid rise followed by an equally rapid fall?

What It Actually Is

If you boil down what this tool does: it’s a daemon running on your machine that has access to everything on your computer, configured through plain text .md files for whatever tasks you need.