You could probably combat this somewhat with a skill that references to examples of the code you don't want and the code you do. And then each time you tell it to correct the code you ask it to put that example into the references.
You then tell your agent to always run that skill prior to moving on. If the examples are pattern matchable you can even have the agent write custom lints if your linter supports extension or even write a poor man’s linter using ast-grep.
I usually have a second session running that is mainly there to audit the code and help me add and adjust skills while I keep the main session on the task of working on the feature. I've found this far easier to stay engaged than context switching between unrelated tasks.
I can totally understand where you are coming from with this comment. It does feel a bit frustrating that people are rediscovering things that were written in books 30/40/50 years ago.
However, I think this is awesome for the industry. People are rediscovering basic things, but if they didn't know about the existing literature this is a perfect opportunity to refer them to it. And if they were aware, but maybe not practicing it, this is a great time for the ideas to be reinforced.
A lot of people, myself included, never really understand which practices are important or not until we were forced to work on a system that was most definitely not written with any good practices in mind.
My current view of agentic coding is that it's forcing an entire generation of devs to learn software project management or drowning under the mountain of debt an LLM can produce. Previously it took much longer to feel the weight of bad decisions in a project but an LLM allows you to speed-run this process in a few weeks or months.
All those small micro decisions, discussions, and dead ends can be recorded and captured by the AI. If you do something that doesn’t make sense given past choices, it can ask you.
Gradually, over time, it can gather more and more data that only lives in your brain at the time you’re building. It’s only partially captured by git commits but mostly lost to time.
Now, when you change code, the system can say, “Jim wrote that 5 years ago for this reason. Is the reason not valid anymore?”. You might get this on a good code review, but probably not. And definitely not if Jim left 2 years ago.
The three most common things I think about when coding are DAGs, State Machines and parsing. The latter two come up all the time in regexps which I probably write at least once a day, and I’m always thinking about state transitions and dependencies.
A lot of the time they are selling themselves as influencers on the subject. It’s often a way to get views or attention that they can use in the future.
I love having generated code in the same repo as the generator because with every commit I can regenerate the code and compare it to make sure it stays in sync. Then it forms something similar to a golden tests where if something unexpected changes it gets noticed on review.
I wonder if the more we use LLMs the more our written patterns will begin to match them. Those of us who work with them the most are likely to be affected earliest.
I don’t know too much about about the TS ecosystem but do these new systems you talk about do it via a “Smart Constructor”? That is the pattern I typically use to solve this problem in Haskell and Rust code.
I once found a confirmed bug in SQLite one time, and it’s the highlight of all my OSS bug reports.
Interestingly enough I found it while fuzz testing a query builder and cross testing it against PostgreSQL, MySQL and an in-memory filtering engine I wrote.
Has anyone ever implemented a schema where you define a view to return the latest instance of each resource, and it maps INSERT, UPDATE and DELETE to an append only table underneath?
I’ve done some of this with an event sourcing system, but not the query rewriting. All my reads use the “latest view”, but the app uses functions to write to the underlying scheme rather than doing “simple” DML.
I wondered if keeping the storage as append only under the hood would keep the app layer simple.
Its usually leadership who are the problem if a whole department is failing. If they started firing from the top down they'd eventually reach competent people who stayed with the department out of loyalty or because they felt responsibile to hold things together. Seems such a waste to fire those people too.
I think the point the GP message is making is that, prior to review/merge you extract atomic commits from your WIP that tell a clear, concise story of how the change was made. The reviewer has less built up context so by chunking it like this they can step through each commit one at a time.
IMHO the expectation is that each commit would 100% pass CI, so if you decided to extract some commits and merge that early you can. This is especially useful when a 6 commit PR is reviewed, and the first 3 commits are fine but there is more feedback on the last three. The reviewer can split the first 3 good ones out, get them merged and whittle down the PR to the remaining three. The subsequent follow up will be less.
IME team velocity goes up with this too, and it encourages small and easy to review commits like a Remove to be extracted and merged early.
Since PRs are always as large or larger than commits, I would much rather have a specific commit flagged than have to wade through the whole PR diff. If the PR is not familiar to me, I want to increase my effectiveness narrowing down the cause, so I can fix it faster.
Also, why don't they share the lists with each other? They don't need government help to do this.
If there are privacy issues then they could just share a list of hashes and if someone's legal name hashes to something on the list then you deny them.
I just hope the LLMs don’t come for parenthesis for aside comments.