We should be more tired than the model
vickiboykis.com
vickiboykis.com
Prompts like "move the code relating to SQL query analysis into a new file", "look for opportunities to use pytest parametrize to remove duplication in that test", "rename method X to Y".
Early indications are that this is helping a lot with the problem where it's easy to churn out thousands of lines of code and not really have it stick in my head, even if I review every line of it.
Reviewing code and actively refactoring it is less tedious and more mentally engaging than reviewing code without changes.
If this was a human collaborator I'd be worried that I'm just creating busywork for them, but I don't care about busywork for LLMs!
The goal is to produce code that I understand and that I can remember just well enough that I get an updated mental model to help me productively make future decisions about the codebase.
It’s almost like a buffer space would be useful for code.
I’ve been using tuicr for agent code reviews and have been enjoying that. I think I’ll try your idea as part of my workflow.
There’s a lot of overlap there with the sorts of things traditional automated refactoring tools can do approximately instantly, locally, and for free.
Mature workflows for those kinds of tasks have been mostly ubiquitous across professional-grade engineering tools like those from JetBrains or Visual Studio itself for longee than many people here have even been working in the trade.
It's clearly not the case for simonw, but much of what many people task AI tools to do foe them are only a novelty for the "VS Code"-type users who stubbornly refused to explore more professional-grade paid tools in the past.
Yet for many tasks, those mature paid tools provided reliable and efficient features that make the AI approach look like an expensive, slow, and dangerously nondeterministic regression.
I've never liked the larger IDEs - VS Code only won me over because it was indistinguishable from a lighter text editor at first, and the IDE tools then emerged slowly as I used it.
This has a list, but some of them that I've used extensively:
- Rename anything, can also do smart renames that apply patterns, deal with capitalization differences, pluralization, etc
- Move function, constant to a different file
- Change function signature, including adding, removing, reordering, and renaming arguments
- Turn a chunk of code into a constant, variable, or function, including replacing other similar patterns or calculations with the new constant/variable/function
- Inline any of the above.
- Replace a class with an interface, have the class implement the interface... same with field accesses
- Move functions up/down the type hierarchy
I'm sure there's others. This has all been there for at least a decade.
1. Find the code you want to change
2. Run the tests to confirm that test coverage is good for the starting point
3. Track down everywhere else that might call or interact with that code
4. Update the tests (red/green TDD)
5. Alter the code
6. Update the things that call the code
7. Run the tests again
8. Apply linters/formatters
9. Address any feedback from linters
10. Check to see if any documentation needs updating and do that
11. Land a commit with a descriptive commit message
I can get all of that done with a coding agent with a single sentence prompt - especially if it's already in a session where it knows that I do "red/green TDD".
... and then I can work on something else while the agent is churning through those steps.
I guess the difference may be in people's mode of AI working: Do you primarily develop in your IDE or a bunch of terminals running vim, and occasionally fire up claude to do more complex things? Or do you primarily develop in a long-lasting claude terminal, and occasionally tab over to the IDE to watch/codereview? In other words: What dev tool is on your primary monitor and what's on your secondary monitor? It's getting hard for developers in one camp to discuss coding and see eye-to-eye with developers from the other camp.
There are a lot of small refactorings that I wouldn't consider to be worth 15 minutes of my time, so I wouldn't do them.
Outsourcing those to an agent means I don't have to make that tradeoff, which means I can get better quality code.
But yes, for a lot of my work I'm now a Claude Code / Codex first developer. I run Zed so I can navigate the code and occasionally make small edits.
1 - Find the code
4 - Move the code
10 - Change the documentation
You don't do the other steps because it's deterministic and always correct.
I don't trust any refactors until I've seen the test suite pass.
(OK, sure, "rename method" might be OK, but most of my refactors and design changes are more interesting than that.)
It also takes away the mental tax burden of thinking up good names for things when writing out code quickly on the first pass.
Works very well with PHP (in PHPStorm), and I'd expect somewhat similar reliability in PyCharm on a Python project with some type annotations (or maybe even with none?).
With the advent of LSPs and treesitter parser approach I wished to see higher level refactorings possible by now, but innertia in tooling space died down as most focus switched to LLMs.
And I sometimes prefer searching for the method names and changing them one by one (find-grip-dired followed by query replace) just to see where all the code is being used, if anything looks bad, if there is some opportunity for code reduction, etc. not that I don’t also prompt Claude to rename things, or better yet make sure all the SQL commands log enough etc.
Anyway, Python tools are usually capable of refactoring any code that isn't reflective or self-rewriting. There are places for those two, but those are very few and the AI won't have any idea how to work at those places either.
JavaScript is a problem.
Even the doFoo to performBar is tedious because you need to catch all instances and your find/replace script strategy might have unintended victims.
In this case indeed, it's just much more convenient.
Even harder would be to update your setters from `void setFoo(Foo f) { this.foo = f; }` to fluent-style `Parent foo(Foo f) { this.foo = f; return this; }` - I'd be surprised if there's an LSP action for that at all. (I'd love to be proven wrong though)
I actually had better luck asking codex to write temporary sed scripts based on the requirements then apply them.
And instantly, certainly when compared to AI.
Ruby would be the one exception I’ve worked on in my head, and for they language, ctrl+f *usually* (but not always) finds the rest.
Ruby is particularly magical with being able to evaluate methods from dynamic strings into running, production code[0]; but otherwise, languages and capable IDEs, like IntelliJ and Visual Studio just support that. I don’t happen to use VS Code, but I assume it has basic refactor, too.
[0] Devise. https://github.com/heartcombo/devise/blob/main/lib/devise/co...
Like you said, it's basically instant.
It's not unlike some managers who tell their teams to do something trivially easy that they could have done themselves.
(I'm not saying this is ideal and I'm not defending my laziness. It's just the current state of things.)
As an engineer: of course not, keyboards are more efficient.
Also as an engineer: of course, “tea, Earl Grey, hot”, or what are we even working for?
My son loves the always ready hot water.
Because it's often not just Ctrl-C/Ctrl-V. It's Ctrl+C/Ctrl+V, except today paste is Shift+Insert, and in the other project it's usually C-y, and you actually need to check one more place for full list of things to copy, and then you discover one case needs a minor transformation, ...
LLMs can handle such annoying details intelligently.
> What's next? "Claude, rename the function doFoo() to performBar()"?
Yes. Such prompt can be issued in many equivalent ways, and works across environments, contexts and tool stacks. I can issue it from the phone, in form of "also doFoo -> perform... reame", and it will work even on Lisp code inside Word documents it accesses through Google Drive.
And even with small things, it's just toil. You can start a small update and then realize with all the back and forth of edit/test/edit/test that you've been stuck in the OODA loop for 40 minutes you will never get back.
Secondly, no, it's not even comparable with AI basically delaying electricity decarbonization on entire continent-sized countries, requiring entire coal-fired power stations and stealing all the ram and storage of the planet to build specific and short-lived hardware that will last few years and will be thrown away very fast.
Getting them to run ast-grep is really fun, especially when it saves me from having to memorize that syntax myself.
any recommendations?
This is technically true, but lets not act like we haven't seen immense improvement of both models are harnesses for these models in the past years. They may not be learning, but they are getting better
As a recent example, I recently had to abandon the multiple LLM reviewer/verifier model I was using because zig 0.16 was released with major changes.
I actually reverted back to full self hosted because the foundation models we’re trying too hard to revert to the older versions of the language.
It is going to be a balancing act and there is fundamentally no way for LLMs to get around this.
We will have to develop methods to do so, most likely by focusing agents on problems that are more static.
If you have zig installed, you can run ‘zig std’ to see that.
You still have the limitations of attention etc…
Even zed’s agent will leverage that built in tarball, but it doesn’t solve the problem, especially as some of the languages killer features are unavailable in C and other languages.
Getting the agent to grep std, example code, comments that reference inaccessible security or bugs etc.. help a little.
But for my needs, not refactoring would just be stepping over dollars to pick up pennies.
But yes it is a problem.
Today's models often output working code. I've had OpenClaw instances one shot simple static web-page HTML, Apache installation, and deployment. It may not meet modern standards or be as secure as you'd like, but fundamentally this is an improvement from previous models.
lmao I hope I never use your products with anything sensitive ever
I don't go along with their mitigations though.
In programming we have one tool for this: abstraction. Decomposition, pattern recognition, even data structures and algorithms are all down stream of abstraction. Collectively, we've never truly mastered abstraction, but it's what we have and we collectively wield it well enough that it's usually somewhat effective.
We are in dire need of a better abstraction.
English to program is the same kind of abstraction as going from assembly to program to C to program to high level scripting to "4G languages" and so on, and hiding all kinds of details behind a much terser layer. It's just that it's a qualitative jump at it.
Here's a ChatGPT provided answer asked to "define abstraction in programming, like when going from assembly to C to scripting, etc":
"Abstraction in programming is the process of building layers where each layer hides the details of the layer below".
Abstraction as a layering idea without regard to how it works is like the pop-pych version in that it is "right" but misses nuance.
And I'm not sure how this relates to TFA's point. Are you saying we collectively need to get better at abstraction so that LLMs get better at abstraction (either by training, or our prompting), so that their code is easier to read?
> Are you saying we collectively need to get better at abstraction so that LLMs get better at abstraction (either by training, or our prompting), so that their code is easier to read?
No - our current abstraction for coding agents is a loop where we express some freeform specification of a goal, then a sub loop kicks off where an llm takes a stab at what good looks like for the next step (make an edit, search for info, run a command to cause some side effect etc etc), it iterates in this loop and when it's finished its sub loop, it declares end of turn and the loop returns to the user for steering input.
That inner agent loop can make it quite hard to stay in control.
What if instead of only these low level free form prompts we additionally had some higher level primitives to work with?
I understand the rationale behind this, but can't help feeling that this is a downward spiral. The software industry has always been a hard place to build and sustain a career because of the pace of change. With these tools, the pressure to increase output is going to grow, jobs are going to be axed - so software devs need to work harder to stay relevant. Weren't these tools supposed to make our lives easier?!
But of course AI is also making union/worker pressure matter even less, since it's function is to cheapen the cost/leverage of workers.
So the only solution is fighting that at the political/legal/social level. Which I ain't see happening anytime soon.
Maybe, but to eat it you need to kill it first or it will stomp you. And this can only happen all at once :)
If there's this huge productivity boost what is it being spend on? I know, many have been laid off, but that's not universally true. So we have a productivity boost that doesn't really deliver anything and overall quality a lot of products/code/writing/communication is going down, yet we spend an ungodly amount of money on datacenters... for what, just spinning the wheels?
In a late stage capitalism market economy, their only actual requirements were to make profit for the shareholders and VCs.
If that means making our lives harder, firing most of us, making us stupider, being addictive, being used for surveillance to sell us shit or control us, or even being used to kill people, all of those are fine, if they fulfil that requirement.
I find this surprisingly useful. The quiz forces me to put in effort in thinking through the problem and solutions. And this effort likely helps in learning, understanding and retention. I also find frontier LLMs are very good at this type of Socratic quiz; they give a very good semblance of having a "theory of mind".
I made a Socratic Quiz skill as part of suite of code-agent productivity tools:
https://pchalasani.github.io/claude-code-tools/plugins-detai...
I am not yet convinced that the skill loss is as bad as people make it out to be, nor am I convinced that "taste" degrades like "skill" does. After all, it's easy to teach someone a framework. It's very hard to communicate why something is not as functional or why the UX isn't as good as it could be.
All this to say, I believe these are valuable questions, but we should be careful what presuppositions we bring into the equation. My value has never been my ability to write code. It's been my ability to make people's problems elegantly go away. I do not believe AI is making that worse (yet).
Side-note: If it IS making my skills worse... I'll just sharpen them later. It's not like experimenting with AI for a few years is going to permanently disable me.
Years of production experience, wisdom, and using something in anger matters for both skill and taste.
We have chatbots in a sidebar that will just generate code for you or, more helpfully, answer your questions. We also have inline LLM code completion, which I've turned off completely because they're incredibly noisy.
What I want is something between those. My ideal use of LLMs while coding would be, i start writing a function and need to act on some data. I don't know what method to use, maybe I'm in an unfamiliar language/framework and don't know what my options are. I want the AI to explain what methods I can call to do X in this specific place, no more, no less. It would need to know what outcome I want, which would be hard to do without jumping out of the code and typing into the chat, but I basically want it to function like Intellisense on steroids. Something that doesn't break my focus.
Current LLMs are anti-flow. For me, that's poison.
I agree the tools in general a big problem with current generation AI tools but there are options if you look around enough. Though you can burn a lot of time figuring out which ones are useful.
I've found that it's much more rewarding to use LLMs as an aid to deep work instead of a substitute for it, and it's even helped me feel more optimistic about my place in this field after a couple of days of getting used to the mental friction again.
I like to do the opposite, asking the LLM to give me relevant follow-up documentation, like the actually docs, where I can read and understand things myself. Data structures, techniques, etc. I still like to read that from the authors, much easier and trustworthy to grasp.
For what it's worth, I still stand by that strategy 30 years later. Delegation is one of the most powerful tools we can use, and developers are usually really bad at it. I speculate that this is because we have egos and some part of my reptile brain feels like bringing in people who are better at something than I am is some kind of intellectual defeat. In reality, the opposite is true.
The lens that folks [who are anxious about forgetting how to program when using LLMs] should use is to ask themselves if they have the same existential crisis when hiring someone or working as part of a team. Does delegation make you dumber, smarter or something else?
In my case, I'm rarely as smart as when I effectively delegate things I'm less good at. In a universe where time is infinite, I could perhaps decide that it makes sense to be good at everything. In this reality, I'm more aware of how finite time is every morning when I wake up. It makes no sense for me to learn C++ for the few moments when arguably it would be handy if I knew it.
Ironically, it was ultimately video games - specifically, some hobby Unity development - that led to me learning just enough C++ to embarrass myself. And my dad never approved of video games, either.
Anyhow: if hiring someone to work on parts of your project doesn't make you less good at programming, I highly doubt that using an LLM makes you less good at programming either. In my experience, I'm actually feeling like a better developer these days because I'm getting so much more done in a shorter amount of time.
Now, I'm still trying to figure out what I can fully delegate to an agent and what I can't. Right now, LLM feel like a senior on technical stuff and a junior on decision/taste. One thing is sure, I can't delegate a critical development to it, not without review. For that review, I need programming skills. Maybe in the future that won't be the case, but I am not seeing that right now. (Using Claude code)
One thing I've noticed in my own behaviour is that the more tired I am, later in the day, the less rigorous I am about auditing what Opus suggests for me. This specific detail is a blind spot for many, I suspect.
The problem is always going to be the 2% of the time it does something horribly wrong on an architectural scale. You don't know when the poison pill might come.
Conclusion: if you wouldn't drive after drinking, smoking weed or staying awake too long, perhaps you shouldn't commit code generated by an LLM that is really amazing 98% of the time.
Tutor-mode is a SKILL I made to have the LLM guide and teach through questions and hints.
So I delegate completely the boilerplate/build system/CI and switch to manual+aided by AI for specific implementation part of the code.
I also keep a final validation step to me because I've experimented a few too many "sorry I told you I passed the test suite and I did, but the build failed and I didn't told you".
I think this is how we should be reading code as well.
First understand the top level. Then the next level of detail and so on. I treat my understanding as graph of interconnected black boxes. If I don't understand a particular black box or a node in the graph. I click expand on it, grok the details and then collapse the node. Here's the grokking details of a particular sub-node also follows the same structure as understanding the root node. You don't need to understand everything from the get-go, expand your understanding on the need-to-know basis.
I'm really trying to do this too. The problem is it's *so easy* to let your standards slip, even for just a moment, and that piece of code suddenly becomes foreign.
I find more mental energy is spent on restraint than execution these days.
By which I mean, we should -- as software engineers -- be insisting on tools that put us in the driver's seat more.
Instead we're letting the agent drive. (I'm as guilty as this as anybody). But really we're letting Dario, Sam, Boris, etc. drive. And it should be clear from their public pronouncements and emissions that they don't have the best interests of our profession -- or the quality of software engineering generally -- in mind.
Yes, certainly, alter how you use the tools. But we need to fix the tools themselves.
And the company says "fuck you then, I'll fire you, and keep fewer coders, willing to keep the AI dance".
The role of AI as a tech is to put you out of the drivers seat as much as possible. All the way to job elimination.
Solution?
And those execs will get their bonuses anyway, and will be drinking their champagne far away from their executive roles and the company by the time that's felt.
There is no mitigation for that situation which would allow human coders to retain their place in the ecosystem.
A man can dig a ditch in a week for £500. 7 men can dig a ditch in a day for £700. 1 man and a digger can dig a ditch in a day for £250.
"But I use knowledge and my brain to do my job, my job is special!"
Not any more it isn't...
They're crude, and also implemented as a sidecar to the actual coding process. Just "hey go do this" which.. I mean... fine, it works, but it's not exactly helping with knowledge acquisition and maintenance.
To make the work easier for humans to do.
What is the point of knowledge acquisition by humans and easy maintenance if an AI is doing the work?
Some people paint, but if you want to record a scene instantly you press a button on a camera. You don't need to know what the camera does to record the image, you don't need to know how to paint, or anything to do with image manipulation. The tool takes care of that for you. Humans have always offloaded work to machines any chance they get. Other humans have always cried that those tools are the end of human ability, from just writing in 370BC to the smartphone in 2007.
Second this. This is why zuckerberg is dying to spend as much as he can to make meta an AI company.
Being conscious of what type of memory you're working in (or need to engage) may be the trick to building rhythm or flow, or whatever. Depending on the case the LLM may not even be necessary. Use something else.
The trap could be in trying to depend on and work with a model the same way we would work by ourselves, as the author describes, letting every type of memory unconsciously operate.
I hope the field moves out of the TUI with prompt + pull the lever paradigm soon‚ when it comes to agentic programming. And the Markdown paradigm too, tbh.
There hasn't been anything that really sticks yet for a shift to happen.
Now the question to the round: in your opinion, are LLMs ok to learn in this way? At least on the theoretical side of things?
One thing that I look at is pushback rate: what percentage of the agent's proposals are rejected or critiqued? If it's below 5% I have found I have gotten too credulous and I am no longer closely following. Danger! If it's above 50%, I have clearly not given the agents sufficient context to perform the task and need to update my harness and instructions.
Who watches the watchers? I can imagine a guard dog process that halts the session to yell at the human if it detects complacency: if the human is providing too few tokens per minute of new context relevant to the task.
As a dev with 40+ years behind me, I apply the same skills and tools to LLMs as I would to a small team of junior devs, except with a higher degree of rigor. I don’t feel that my skills are atrophying, in fact I’ve learned a bunch of new stuff in the process.
Depending on the work, I’m producing about 4x what I would have with a small team, and it is better code, better documented, more modularity, and less bugs than I was shipping with humans.
You have to keep a tight OODA loop, and write the code in specifications, protocols, interfaces, and implementation plans before the first executable line is put down. Iterate the code and docs together to minimize and document drift.
Treat a context window as a session. When it’s full, you start over with a new session, bring in the onboarding documentation, and go to work. I never cross context boundaries midstream, it’s always a disaster. 1m context is critical, I burn 150k just onboarding and orienting a new session.
Still shipping 4x at 1/10 the cost, and less bugs in the field (greenfield firmware development)
This made me chuckle a bit.
There’s no point in fooling ourselves about our own skill retention if this is the case.
That being said, I have deadlines I have to meet.
Imho most of professional coders trade their time for oblivion.
The problem is that you really don’t remember anything about the code. It is not your creation.
It’s like a monkey in front of a slot machine, just pulling the lever and waiting to see if it hits the jackpot.
At the end of the day, it remembers that it pulled the lever. And how many times it won :)
Agentic-based coding with /goal and multiple agents coding together is another level…
But the issue remain imho - if there is an error, who is going to repair it?
The next problem is few care about that, at any level: coders, managers, execs. Just want their feature churn.
The even worse problem (or maybe, a positive) is that most of that code and the products powered by it aren't needed either.
Clankers are for reading code, not for writing it.
I mean you can, sure, and doing so will produce a journey filled with experiences, but no business would operate like that and stay in business.
That is the future of jobs where thinking is the work.
I typically prompt the AI with a detailed plaintext description of the algorithm and have a firm grasp on the time complexity. Perhaps my low level implementation skills are atrophying, but now I have greater bandwidth to think about the higher level goals now that I don't need to worry about the nitty gritty stuff.
It's kinda like how we went from handwriting machine code to programming in FORTRAN, and then going to a higher level language like Python. To me, using AI is just another level higher. Skill retention of understanding machine code and assembly went away in the majority of software engineers, but we gained a lot more than we lost.
Thanks for this insight and validation
The experience will feel uncannily similar to AI generated code. So treat slop the same way. Give it a good, well tested API, and file an issue or PR when something breaks.
Can we be 10x more productive though? Or is it more like 1.25x? Is it AGI or is it more like an advanced compiler?
Unfortunately the world is betting on 10x when the reality on the ground feels more like 1.25x.