Morning bath robe rant: it’s a good thing.

I'll keep insisting! 😂 This keeps me in the loop, gets me documentation for free, and includes tests. I haven't found the case where the test is neutered by the LLM just so it passes, but now I have anxiety about that.

We already do this on embedded software dev and yes it gets complicated. You have E2E timing IO testing suites, json fifo event logs, critical logs, etc. Anyone who thinks that general AI LLMs will simply take over all of this is fooling themselves. Future is specialization.

Tests are the chains with which we must bind this giant if we would make him serve us

Shared with my friends and colleagues.

explain this at "token management".... Even the dog understood...

@GeoffreyHuntley Couldn’t help myself to do a /code-simplify on my build #ralph loop before commit.

I vibe-coded a quick tool using WindSurf and forced it to Red / Green / Refactor; what it came up with was better than any Leroy Jenkins vibe-coding I've done.

I was vibecoding a pretty big app with AI for 10 days straight, and reached the point where it just could not comprehend the amount of code it wrote, so I had to ditch an AI for 2 days straight and fix everything manually. Now then I’m done, I can say that for two weeks of work it took to implement every requirement, couple years ago it would take me couple months at least, if I’m not overestimating myself.

Close your eyes and picture Agent Smith from The Matrix

@ThePrimeagen @thestanduppod @trashh_dev @cmuratori Request to get Uncle Bob on the Standup to talk about this.

Kudos to you, your influence goes beyond humans now

It'll be interesting to see how this plays out because some folks will treat the code as a "black box", assuming AI will deal with it. So the idea of doing things like "keeping the code clean" will only be done by some. Now whether they're the smart ones or whether they're wasting their time will be interesting...

What a focused dog! 😃

My only quibble with the statement is that he is conflating coding with current languages. We will always be coding -- it's just the the language we will code in will be a mix of natural language in both a formal specification and an informal conversation.

Just blink if you are threatened by the dangerous animal.

@unclebobmartin are you starting with high level spec and some acceptance tests you wrote yourself or having LLM generate acceptance tests, or some combination thereof?

I started with a high level spec. Claude and I collaborated on an initial set of acceptance scenarios derived from that spec. Then I play tested the game and noticed several bugs due to ambiguities in the spec. I instructed Claude to repair those issues and write scenarios for them. I reviewed those scenarios. Then I decided on a few new features to add. I instructed Claude to write scenarios for those features, and after reviewing them I allowed Claude to implement them. I added a rule that no new features nor significant changes can be added until scenarios are written.

I think you've experienced the golden era of software programming. Lucky. Now we in the post-golden era. We need luck.

I think the next era will be better than the last. Once all the turmoil has subsided (a matter of months) businesses will realize that they still need highly skilled individuals to wrestle these cantankerous machines to deliver value. The more things change, the more they remain the same.

Too many new things to learn not sure how it will shape, we can just try our best and see how it unfolds.

Roast me if you find this dumb BUT my current best strategy of context compression is having a root level docs folder where I have an md file for how every larger system works at a high level. Does this violate the single source of truth principle? YES. Does the AI not waste context scanning every single file or unnecessary files when working on these systems? Also YES Instead of scanning 10 files (maybe 3000 lines of code) it will scan 1 md maybe 300 lines of code and maybe sanity check a couple files. Thats a simplistic example. Caveat - you do need to audit your docs folder for updates But this definitely has improved the life span of my agents memory and prolonged higher quality output I find this to best reflect how I hold multiple systems in my head - core principles and data flows there but the details missing.

Awesome dog. I have a chihuahua shih tzu mix. Small dog love.

Your name and clean code is in my Agents .md file 👑

It's crazy that just asking the AI to create unit tests leads to better code.

@GeoffreyHuntley genuinely loving this uncle bob arc

I think I’ll keep my focus on hand rolling my own code. Something to think about though! Just because I make it by hand doesn’t mean I can’t make it readable for people and machines.

The core of the bad news: "Big-AI isn't currently addressing provenance in a meaningfull way, and given Sam's, Elon's and Dario's recent statements, it might never." Some of my ideas about "taming" this and other bad news that might help with insights:

I love that you dove headfirst into this stuff and pushed yourself. Too many people I know are poo pooing agentic coding, not realizing it could be to their detriment one day.

you are a genius i completey agree im exploring a dimension of this with guardrails and evals as a part of this acceptance testing

Thanks 🙏🏾

I’ve been thinking that in this future, the functional paradigm will really excel, favouring simple and decoupled architectures. Easier to write good tests for, and they also do really well on token efficiency (Clojure especially). Does this match your intuition?

Thanks for the advice! I'm quite sure that's the right position to take as a developer en 2026+

This needs to be included in a new book for the new era: Clean AI. I’m waiting.

Hey, @unclebobmartin agreed, and I would like to suggest that 1) Use TDD for new feature implementations. If the agent generates the test, the engineer must review it before the agent continues building out the new functionality. 2) Agent must use unit/e2e/acceptance test as the success criteria. These tests must be in place before modifying existing code. If you always follow TDD, then the former comes for free. Tests become the source of truth in this agent era

Wise words

It's up for the humans to set the context of what the job to be done is and for whom, draw up the contract of what the agent for the AI needs to accomplish in that context, and then draft a constitution to govern behavior throughout the whole process. Without these three, GFL.

@unclebobmartin I wonder if you take out code and only leave the existing acceptance tests and unit tests, would the model generate the app/game correctly? What would happen if you would do the same with different model? Those would be very interesting findings...