Demo the work, keep changing. Don't worry about the amount of changes, don't worry about nice commits. You are free to explore and produce as much changes as required.
Then use AI agents to clean up the history and produce a nice change set that you can ship in a more traditional way, small clean commits.
That is an interesting idea I guess. AI agents can be used to clean up git histories, not only produce large amounts of code.
Somehow, I think heavy LLM users don't really care about a clean git history / changesets? May be if the project or existing workflow you are bound to requires it you will do the extra steps but really I don't see it being done much.
Are you really going to review these small changes? So far the example I saw from agents are terrible. Full of drive-by, unrelated changes here and there. Mostly harmless or OK changes, but not focused.
Some do, some don't. Yes, it depends on a lot of what wider organisation you are trying to fit into.
But LLMs can traverse that tree at will in seconds and mine valuable info. Commits description value in my experience has shot up dramatically _with_ heavy LLM usage - they are forgetful and commit messages are just the right place to keep “context that might be needed, but not worth adding to the documentation corpus” - they are the “how did we get here” to the “current state” of docs files and code comments.
It’s worse than that IMO. In over 10 years I’ve yet to have seen a team that even knows how to rebase, asking them to investigate commit logs or bisecting would be like teaching a cat to bark. Insisting on “good commits” (or even worse, conventional commits which is a poor solution to the wrong problem) is just adding bureaucracy where it’s least needed.
I think it’s like with code formatters - now we have a tool that does the tedious thing on its own and unlocks human creativity where it is warranted… at least in the ideal case.
What I'm getting from tfa is that the point is to fix exactly this. Changes get packaged into focused changesets that make sense together. I've done this manually when working on a hairy feature, not splitting it into multiple pull requests but rebasing the history into (ideally) atomic commits that make sense sequentially.
We generate an AI branch immediately every time a spec is completed. It then gets a brief review in our team meetings when the spec is introduced. We just take a few minutes to walk through what the AI did and take notes on what was insane and what was a good approach.
#1, this informs us about the actual quality of the spec, if the model did something wrong it will often become evident that the reason was an omission, ambiguity or contradiction in the spec.
#2, it saves the developer a lot of time, because in ~3 minutes we've usually established that at least half the code the model wrote was acceptable for use. Dev tasks look more like "Here's the implementation in concept, but pieces X, Y and Z will require your expertise to formulate a cleaner approach." And since AI tends to do a cleaner job at smaller scoped problems, it tends to be a great help to the dev at this stage too.
If anything I feel this process increases quality, not reduces it. It certainly speeds things up. Most major challenges and problems are immediately exposed by the model's proposed implementation. We don't have to wait for the developer's first cut to discover them. They go into the task knowing exactly what to focus on, and having the uncontroversial parts of their eventual PR essentially "pre-approved." The result is where we would have had to do multiple iterations of a feature, they often come back at the very next meeting having totally nailed it.
Even before LLMs, most teams don't bother cleaning up. The best you'll see is feature branch getting squashed on merge.
AI is not really a factor here. Teams/orgs that did not care before still won't, only 10x more.
Seems to be the standard in shops with competent git users and no culure established around other version controll systems. What the commit messages are is fluid. My messy personal ones tend to be "start/end of day commit."
It was LLM written or heavily LLM processed. This, for example, is 100% LLM:
> Deciding boundaries up front was never about making the build easier. It was about making review possible, and it was the only affordable way to get there. That's the part that changed.
Producing massive changes in a codebase has never been impressive. LoC has always been a bad measure of progress, and yet, modern frameworks, popular tooling and haphazard business logic almost dictates that you blurt out as many LoC as possible. Software engineering wasn't supposed to be this verbose.
It's amazing how the AI can learn from branches that ran up against a wall.
I don't think that you need to make wild iterations on the real product, just do it scratch space and 'merge the lessons not the code'
This whole 'design up front' never worked well because it baked in to many bad assumptions.
The experimental efforts help you work through those.
Once you've worked through it - toss them aside and then let the AI plan based on the scratch work.
The tangled stuff is in scratch space, that's what it's for.
And critically: it costs almost nothing.
The power of AI is not the 'nits and small bugs' - it's the firehose of scratch branches it can let loose in where it doesn't have to worry about PRs, reviews.
Once it's all figured out, then the path to the solution is actually relatively clear - and it can be 'planned out' and just done.
> Once it's all figured out, then the path to the solution is actually relatively clear - and it can be 'planned out' and just done.
it is important to be clear about what this buys: quick coding and testing primarily.
the mental effort needed to design, manage the constraints and all other essential workflows to bring about a professional product is still present.
What it buys is a clear path through the mess without having to make it.
I'm saying: 'figure out the design and constraints by discovering them ... instead of trying to 'one-shot' them as a human'.
By the time you've run the gambit 100 times - you can synthesize the learning into something that's nearly complete.
Let the AI figure out the constraints, you just guide that. It can't do it in conversation, it takes a lot of iterative prodding.
But yes you have to decide where to steer it, and this only works for certain kinds of engineering at a certain scale.
I got fired back in June for protesting against trying to “design everything up front” (by adding more slop plan docs alongside the hundred of others that my tiny team was charged with producing per month) and instead opting to explore the problem space through code/experimentation. In actuality, I just asked to be remote again. I was burnt out and they were looking for an excuse to fire me.
Authors: If you actually put deep thought into the content and then have an AI do the final writing, it immensely devalues your final product. I think this author actually has some good content here but the AI-generated prose does it no favors.
Companies: If your blog is for recruiting purposes (which is often one major motivation), AI slop articles are going to be a big turn off for the best talent.
Great round up of the meat and potatoes of engineering in a mature codebase without giving in to entropy.