Morning Bathrobe Rant: AI slop.

Spending 300% of the effort to prevent AI from writing slop, instead of just writing the code yourself at 100%

You are badly misjudging the comparative effort.

I’ve started doing this for the first time in my life because I don’t want a guy with a stare that angry to be mad at me.

Core piece of C.R.A.P. here - a quick way to ensure that unit tests are effective and you're not wasting time by applying 80%+ coverage to everything.

I cannot discern a fiscally rational reason to pursue this path. In essence, the AI must write the code, the tests, the performance scripts and every other harness under the sun just to avoid slop. The costs for this are becoming astronomical. I think I’d rather just write the software myself.

You do you. I can go a lot faster, using an AI, even with all the tools I employ.

My question isn’t about speed. It’s about rising costs. How do you honestly expect junior engineers to fork over a king’s ransom to do what you’re prescribing?

So far all these tools haven’t cost me any more than $200 a month. If that multiplies my productivity by a factor of two, then that’s a pretty big win. I would expect my employer to cover a meager $200 a month.

@wmksidebottom Certainly you burnt several thousand USD in compute

@unclebobmartin Exactly. What’s 200 a month today will become a second mortgage tomorrow.

@unclebobmartin you've got this idea you talk about where ASM gave us a 10x productivity boost over machine code, C gave a 3x boost over ASM, Java a 1.3x boost over C, and so on. What's your assessment of the boost factor from AI used right, as compared to the tools of 2022?

EXACTLY! People literally are allowing AI to lead the way when THEY are the ones who are supposed to be using the AI as a tool, meaning the user is the leader! You literally have to tell it what to do and tell it not to write BS. Correct it when it goes wrong. I love your page!

writing 1000 lines of harsh instructions in markdown files isn't enough?

That's what software engineering is morphing into now, I think more than ever, we need good engineers!

So, according to this post: In that past, as Software Developers we were 1. thinking about how solve problems 2. thinking about how to design software But, now we are all just: ===> Baby-sitters for AI. 🤓

It seems like coding normal way is just 5 times easier. All i really wanted is a boiler plate generator.

As it turns out. People created slop too. A whole industry rose up to teach people/teams ways to combat that slop. Yet, people feel helpless to AI slop. Funny how that works.

Enjoy

Modularize the system so the AI operates within clearly defined module boundaries. Use pure, IO-free modules wherever possible to enable fast, reliable testing, and push all IO concerns to edge modules. Test module boundaries through explicit interfaces. You can't fail with this way of working.

What about people claiming to run 20 agents writing code to the same codebase simultaneously? Would it be possible to prevent slop from 20 different agents or this would be a fool’s errand?

I just add “don’t write slop” in my prompts.

But ... but ... but ... adding all that context, contracts, and constitutional governance sounds like work!

You don’t mention the using of hooks (claude code & gemini for exemple) to enforce linters at each edit/write - which I find super efficient to enforce coding standards I was possible to literally force Haiku into respecting Clean Code with no prose instruction at all this way

mutation testing sounds to me like a restatement for rule 1 of the 3 TDD rules. “You must write a failing test before you write any production code.” Is that so?

An expression can have many failure modes. The first rule only demands that you test one of those failure modes. Mutation testing finds all the others. If you are being very diligent, while practicing the three laws, you would note down some of those extra failure modes and write tests to cover them later. But that’s fairly haphazard. It’s very easy to miss subtle failure modes.

Right. I believe the model has to do the mutation testing manually right? Are there any skills/tools you know of? or the model does that out of the box? Is that part of the agentic discipline series? Thanks.

There are several mutation testing tools that you can find. I suggest you get the AI to look for some of them and then have it build your own mutation tester tuned to your project. The model will do the mutation testing manually, but it’s unreliable. I wouldn’t trust it.

Okay, thanks.

He called me out 🤣. @cruelhandeth

Everyone, listen to your uncle Bob’s wisdom! “constitutional AI” still means a human has to write the constitution. My constitution is implemented as a cursorrules file. *I* am to blame if my constitution is broken, and I get slop-leak.

So in the end, the focus shifts to planning and spec refining and less on generating boilerplate. That's not the kind of developer obsolescence future Dario is predicting.

uncle is very wise but looks a little scary 👉👈

I literally look forward to this morning rants. Lol

Throughout human history CRAAP has been the top mass production, did slop finally outperform and take first place?

😂 Hard agree! I've been saying this for a long time. The unfortunate part is how so many people don't want to manage, work, or pay attention. I don't know what it is, but I also saw that work ethic start before AI.

A tool is only as good as its user!

Bob, help me understanding the mutation test part that you emphasized! I was thinking the opposite, always mutate your test because AI will change to make this code pass. I understand that my criteria acceptance is my key barrier here but Its a risk here, dont?

You tell them, Uncle!

Never been easier to get full coverage and great adversarial coverage. No excuse anymore.

Meanwhile, me coding with AI:

Why are you drinking ai slop?

Sir, when can I take a shower? Claude says it's unnecessary.

Nha my AI does produce high quality output because I told the one agent I trust the most to quality check the other agents I don't.

true for sloppy use. but with tight specs, TDD, and review loops, ai can accelerate real craftsmanship the key is staying in the driver's seat