Things that have a bigger impact than coding assistants | Hacker News Reader

Things that have a bigger impact than coding assistants

codemanship.wordpress.com

110 pointsflail59 comments

If you're offended that the author isn't claiming that AI is the biggest productivity gain of the last N years, consider that perhaps you could do these 5 things and still use your AI tools.

The article is about the effect of these tools on an organization. If your org isn't doing these 5 things and thinks "adding AI" will finally make them more productive than ever... they might see modest gains but the article is claiming it won't be as big as if they'd used these practices.

It's hard to measure which has more impact: changing your management style and organization structure or using AI.

I'm willing to bet they both have some impact. From experience I believe the former has a bigger impact. But I'm not sure it's true industry-wide.

Have done some of these recently - Smaller teams are better value/$ spent [Confirmed]

- More frequent releases accelerate learning what has real value [No improvement]

- Limiting work in progress, solving one problem at a time,increases delivery throughput [Continued]

- Cross-functional teams experience fewer bottlenecks and blockers than specialised teams [Confirmed]

Empowered, self-organising teams spend less time waiting for decisions and more time getting sh*t done [Confirmed]

Additionally, smaller teams 1-3 engineers per project who are empowered are much happier. Side effect was time spent on process, tickets, communication dropped dramatically. Time spent on creating and confirming increased.

In a large organization solving your own blockers can be the difference between releasing next week and releasing next quarter. More frequent releases only help in a business where users adopt new features quickly.

Strong title but empty content.

As someone in a work environment that does NOT have these 5 boring things in place, I can certainly speak to how much that slows productivity. So I think there's some value there, but it doesn't go into much detail, nor give you a game plan for implementing those things in your workplace. (And in many cases, you aren't empowered to implement them, so you either attempt to manage up or you find a new job.)

Empty is not the same as succinct & straight to the points. I'd say strong title, succint content. Each of the points invites as much thought as you can give it.

Content is left as an exercise to the readers.

Not totally empty, but would have been much better digging into each of the 5 things and why they aren't commonly implemented.

Author hand waves at "organizations not wanting to open a can of worms" when I wanted to examine each squirmy helminth.

Only because you've heard this advice already.

The point is that these practices imply other practices and propagate their own culture. It's simple and not new but still unreasonably effective.

Any time critics single out “autocomplete” in AI coding tooling, I know they haven’t really played around with this stuff. Autocomplete is barely useful with or without AI. The real game changers are “chat-with-codebase” or agentic development tools (do these have a better name?)

they are "game changers" if you are a mediocre software dev thats churning out crud widgets .

any decent software developer uses and creates abstractions instead of generating reams of code using AI.

From what i've seen at work AI is "game changer" for coding in worst sense. reams and reams of duplicated code that looks slightly different from other generated code doing similar things. Before AI ppl used to stop and create some library now they just generate shit because its so easy. AI is death of software engineering.

Any good engineer enjoys creating abstractions too instead of hoping a machine trained on code golf will solve the problems.

My manager probably adds 10+ hours to my week by pushing Llm code at our projects only to follow up with several merge requests to fix his work. I just approve whatever he pushes because he isn’t interested in actually solving the problem. He’s interested in seeing if he can fiddle the solution out of an Llm. Each time it involves me telling him the answer. His boss is the same way. Literally dragging the company efficiency down and proving the efficiency gains are meaningless.

This is all of junior engineers on my team now. I shudder to think about having to maintain all this generated crap.

You don’t maintain it, you wait until a bug comes your way from their work and pass it down to them to screw around with more. Let them maintain it and eat their own careers.

I'll dump you some of my last queries

> I'd like to output lines where .stack_trace is non empty with JQ

I vaguely remembered that it sometimes has "null" and sometimes has empty strings

Time gained ~60s looking at the doc

> Here is a Jira ticket's description: ```....``` Please rephrase this more clearly and make the text flow better?

^ Then I picked and chose the improvements

No time gained but quality improved

> Critique this for accuracy: (A long comment about the properties of randomness)

No time gained but quality improved

> How do I make excel prompt me to select the goddamn delimiter when I open a CSV file instead of just picking a random fucking one that never works

Question was filtered due to content policy, because apparently they don't want you to offend the robot

not sure how these are examples of

> real game changers are “chat-with-codebase” or agentic development tools

It's not an example of that, just a response to

>From what i've seen at work AI is "game changer" for coding in worst sense

Ah gotcha. yea those are some of the ways i use llms too.

There is a pesudo-mandate at work that all new code must be generated by LLMs so ppl are checking in llm junk without proper vetting.

Why extract common logic into a method when you can just generate junk

When someone throws out a whole article because of a word in a meme image the author used to add some levity to the discussion, I can tell they either didn't read the article, or they just want to discredit the author without addressing the content of the article.

The "article" (the article scarequotes "A.I.", so fair game) is a data-free, extremely low-effort tosser blog post that could have been a tweet. It's pandering to the incredibly boring "ha ha AI is bad! Look, distraction!" head-in-the-sand approach that is far too common.

I use Cursor as an agent and sometimes I use autocomplete. Both tools are dependent on what you are doing at the moment. I like autocomplete when I am focused in tight on one file. I spell things out in depth in file and start out code and hand pick completions. It brings my mind in sharp on what I am doing. Agent is when I am doing big but simple stuff where I am crafting less in detail. Refactoring and setting up tests and basic shallow framework code.

But this article is on point. All of things listed are more impactful than LLMs

I've been using vim forever, and often use "advanced" editing techniques: macros, .-repeat, :g/.../norm ..., occasional templates.

For certain projects I'd used `vscode` (with the vim plugin!), and there's definitely some helpful bits. The biggest helper for me is/was the `F2-rename-symbol` capability. Being able to "rename" securely in the whole file (or function), and across the project is super-useful.

Working with Cursor and the autocomplete is (often) pretty shockingly good. eg: when I go to rename `someVar` to `someOtherVar`, it'll prompt to `<tab>` and:

  * rename the function call

  * edit the log lines

  * rename the return object value

  * ...etc...

In vim, I'd `*` to automatically search for `someVar`, then `cwsomeOtherVar`, (change-word), then `n.n.n.` (next, repeat, etc.)

...so my overhead (by keystrokes) is `*` (search), `cw` (change word), (`n.`) next-and-change. Five "vim" characters, and I mentally get to (or have to) review each change place.

In straight `vscode`, I can do `F2-rename` and that'll get me replace _some_ of the variables (then I still have to rename the log lines, etc).

With Cursor, I make the `cw...` and it's 90%+ accurate in "doing what I probably also want to do" with the single `<tab>` character.

It gets even more intriguing where you'll say `s/foo/fooSorted/` and it automatically inserts the `\*.sort()` call, or changes it to call `this.getFooSorted()` or `this.getSorted( foo )` or whatever.

For "cromulent" code, cursor autocomplete is "faster than vim". For people that can't type that good, or even that can't program that good, it's a freaking god-send. Adding in the `Agent...` capabilities (again, for "cromulent" code)... if you're just guiding it along: "Now, add more tests" => "Now 50% more cowbell!" => "Whoops, that section would be more efficient if you cached stuff outside the loop."

Even then, I have to have some empathy with the AI/Agent coding, "Hey... you messed up that part (btw, I probably would have messed up that part the first time through as well...)". We can't hold them to gold standards that we wouldn't meet either, but treating them as "helpful partners" really reduces the mental burden of typing in EVERY SINGLE CHARACTER by yourself.

I'd describe myself the same as you did in your first paragraph here, and wanted to say that with LSP, your "rename function call" example in vim (well, I switched to neovim some years ago) is (mostly) `<Leader>rn` (default in my LSP setup, I'll paste the relevant line below) when the cursor is on the symbol. That will change the name in all open buffers. Annoying thing is that it doesn't (ofc) change the name in "log lines", so it's helpful to pre-`*` on the symbol name so you can go back and `n.n.n.`. (I've always wanted to "wire up" something to also detect those log lines and change them at the same time, but... ya know.)

The line in my nvim config that sets up the LSP "rename" is:

  vim.api.nvim_buf_set_keymap(0, 'n', '<Leader>rn', '<cmd>lua vim.lsp.buf.rename()<CR>', opts)

I mention this because in spite of using vim/neovim for over 20 years, I still learn new things (from HN comments and elsewhere) -- which is part of why I love it.

(To your larger point -- I concur... btw)

I don't know if you've ever messed with `ctags`, but I used that a long while back with a crufty, large, PERL codebase.

It was super nice to have "jump-to-definition", but the vim plugin in vscode is very nice (missing a few things, but even `<c-w>hjkl` "does the right thing(!) so they're really trying).

I haven't leaped to nvim (yet?), and the fact that vscode kindof "just works" has prevented me from chasing LSP support or setting up the more "advanced" features, but thanks for sharing!

I'd be fantastical if something similar to that cursor autocomplete were available in a "real" cli-vim. There's things that really bug me about the "hover suggestions" (eg: can't always tell which characters/lines are "real" or "suggested", especially with auto-closing double-quotes suggestions), but when I occasionally drop to a terminal vim for "accurate" editing, I really do find myself missing like "I should just be able to tab-complete the rest of these edits..." and I don't know how to express that in an appropriate "editor" context?

Maybe lean on like a `vimdiff` representation, where you could `:vsplit $ASSISTANT` and accept suggested diffs? (Hmmmmm....)

In most cases changing a symbol globally is one find -exec sed incantation, but it's also basic functionality of many IDE:s and LSP:s.

> In vim, I'd `*` to automatically search for `someVar`, then `cwsomeOtherVar`, (change-word), then `n.n.n.` (next, repeat, etc.)

There's also:

  :%s/someVar/someOtherVar/gc

The "c" at the end is for "confirm", vim asks whether to change each one and it's a single "y" or "n" before going to the next.

Maybe. A bigger impact factor imo is the distance between the tech you're using and the center of mass of training data. Showcase site in NextJS with a CSS gradient hero banner and a news letter? AI will be amazing. But then again, so will many no-code solutions. The sweet spot is probably custom enough that you're just outside of Wordpress but within well-trained domains like web and its popular frameworks. If you venture outside the mainstream, quality degrades rapidly.

> If you venture outside the mainstream, quality degrades rapidly.

Whilst I agree, the mainstream is also where the vast majority of software development occurs. CRUD apps and enterprise workloads, etc.

Saying that current LLMs are only useful for the mainstream is saying that they're incredible useful.

> Saying that current LLMs are only useful for the mainstream is saying that they're incredible useful.

Yes and no. If you have a hammer that works 98% of times, you might wanna use it a lot or not at all, depending on the circumstances. IME the double checking misses the point for production code. Might as well write it myself. I’m sure YMMV a lot.

Yup. People have no clue, and it shows not only a skill gap, but a lack of curiosity and initiative to use these tools

Sometimes it's not simply a lack of curiosity, but having the space to build these tools into your workflow. As a solo dev responsible for all development, project management, support, & infra, I've gotten as far as trying to use Aider for doing some features in our legacy codebase but not having it break the time-cost-benefit barrier.

Now I feel like there's probably other workflows out there that I'm ignorant to that could be better, but keeping up feels impossible. Is there a particular approach/tool that you're finding to be really beneficial?

My general flow:

1. Build huge prompt with Repo Prompt (500k+ tokens)

2. Ask for gemini pro to summarize the task using above prompt

3. Hand off coding to codex (openai)

The above flow replaces a junior developer

you seem to be spamming a version of this comment on every AI post. Can you put money where your mouth is.

can you demo this flow on a open source issue on a repo like pytorch ?

I am curious to see how many issues this junior developer can close.

I am spamming it because I cannot believe how well it works. Its really about prompt construction and distilling the code base into something that can be fed to the coding agent

> I cannot believe how well it works

me either. Thats why i asked you for a video but you seem to be more interested in spamming the same comment again. Maybe AI will replace engineers that don't understand what is being asked.

The cynical question that I ask is: Will AI tools make it possible to finish projects? Show me a project that goes from being 3 years late to only 2 years late, thanks to AI tools.

I'm using agents daily and programming is fun again. Adding a small feature in the same time it takes to write the spec is a game changer.

That's good. I like programming and find using agents saps all of the fun out of it for me.

Is this LinkedIn?

What a weak article, maybe should have used a chat assistant to write it. Wait, let me try:

https://chatgpt.com/share/682ddb36-50f4-8004-b54d-3e41a10ab8...

ew smells like slop. Theres enough of that content out there already

Yeah, the original article is as bad as this

This article "DevEX: What Actually Drives Productivity" identifies three main categories of factors that affect productivity: Feedback loops; Flow state; Cognitive Load.

With that in mind we can say that:

> Smaller teams are better value/$ spent

Feedback loop: Reduce the communication channels

Cognitive load: A better defined responsibility area

Flow state: Tasks are more coherent instead of being splited ad infinitum

> More frequent releases accelerate learning what has real value

Feedback loop: self evident

Cognitive load: Easier to justify dropping features.

Flow state: Alignment with the user needs.

> Limiting work in progress – solving one problem at a time – increases delivery throughput

Feedback loop: progression transparency

Cognitive load: self evident

Flow state: self evident

> Cross-functional teams experience fewer bottlenecks and blockers than specialised teams

Feedback loop: Easier to address blockers

Cognitive load: No need to resort to inter-teams communication.

Flow state: Easy to offload part of the problem to close teammates

> Empowered, self-organising teams spend less time waiting for decisions and more time getting sh*t done

Feedback loop: No need to wait for management decisions and policies

Cognitive load: Experiments are easier when there's no red tapes

Flow state: Natural workflow due to the above.

[0] https://cacm.acm.org/practice/devex-what-actually-drives-pro...

> Limiting work in progress – solving one problem at a time – increases delivery throughput

This. Juggling 17 tasks is not a badge of honor. It just shows the world that you have neither the focus to prioritize nor the talent to execute them efficiently. Strange flex, bro.

Put another way, nobody cares about your TODO list, everybody cares about what you've actually shipped. So actually ship things, one at a time.

Ugh ads will be even more weird and annoying. Looks at those faces...

“Smaller teams are better value/$ spent”

Is this true? A more established / larger organization often has higher revenue per employee.

I guess value is subjective but from a pure economic standpoint I think the F100 employee wins?

Nvidia is 2xing a year right now so a smaller team / startup would have to match at least that. Certainly some will but enough to make the point valid?

I think scope likely plays a big role here.

I think there's a misunderstanding. The post doesn't say minimize the number of teams you have, but limit the size of your teams.

Like "two pizza team"s of Amazon.

Your boss disagrees.

Sounds like its written by an AI non believer.

I've been able to write things in a week that would literally take me six moths previously.

The critics don't know the facts about how much time you can really save because their anti AI so they don't use it so they don't know, but they're happy to tell you some "facts" about how your new found AI assisted productivity is less than you really think it is.

When you write things in a week rather than six months, what was the context?

Small teams, more frequent releases, solving problems incrementally, working across the solution rather than having to bridge silos, feeling empowered and safe to try things?

If you already had those things, and took them for granted, then it may be hard for you to estimate the impact they might have had if they weren't present.

Would you still have been able to use AI to build something so much more quickly if your hands were tied by organizational policy?

Sounds like you're trying to find out why I'm wrong about my productivity boost from AI.

I'm not claiming you're wrong about the boost you got.

I'm saying you may have already had 5 boring factors that laid a good groundwork to enable you to get that boost, but you are discounting those factors.

If I'm wrong, then tell me the context - were you in a situation where you didn't have those factors in place?

im interested in the answer because I have not seen the same productivity increases with claude code and copilot. When I read comments like your original one I think that I must be using the tools incorrectly. How do you use them in your workflow?

Or for anyone else reading this, are there resources you have used to learn how to get the most out of LLM coding tools?

> I've been able to write things in a week that would literally take me six moths previously.

you wrote a tic-tac-toe game for android? so what ? what does that prove?

have you seen rails blog demo from 15 yrs ago ?

I think the productivity gains depend a great deal on the tools you use.

Writing Python or Javascript? You're in luck! Wrangling some gnarly SQL or complex Rust? Expect a modest boost. Modifying an old Eiffel program? You're on your own.

I work in WinDbg analyzing crash dumps and I'm trying to get better at it so I've been trying to get Claude and Gemini to help me get to the bottom of some memory dumps I have lying around and the experience hasn't been great.

I used Claude yesterday because I wanted to modify my gvim setup to remember the window size and position and after a few tries I finally got what I want (it really wanted to use mksession but that persists way more than window size and that isn't what I wanted).

Show us.

> Sounds like its written by an AI non believer.

> I've been able to write things in a week that would literally take me six moths previously.

I don't think you get it. For instance: would you have been able to do that if you were on a team that was not empowered, not self-organising, and you were "afraid to fail ... afraid to try – to make calls, to take initiative, to just f-ing do it"? I doubt it.

I'm curious exactly what that thing was that you "[wrote] in a week, that would have literally take[n you] six moths previously."

Also, what species of moth do you use to measure time?

Leaving the supposed gains in programming speed aside for a moment, have you been paid 24x for that 24x productivity? Have you even been paid 1.1x for that week of work?

> Leaving the supposed gains in programming speed aside for a moment, have you been paid 24x for that 24x productivity? Have you even been paid 1.1x for that week of work?

Don't be silly! The passive, hardworking owners are the ones that deserve all of that profit. Those are just the new expectations. He shouldn't be greedy, and should be happy that he just gets to keep his job.

Awesome! Hopefully your workload won't be adjusted up accordingly.

Reply on news.ycombinator.com