We need to keep track of many things happening in parallel and read enormous amounts of generated text. The constant guardrailing, decision making, and context switching overload me.
I am pretty sure most of you are experiencing the same as me, but: Have you found ways of using AI without increasing mental overload?
I would love to hear your tips, but still be relevant in this AI era, so my boss does not fire me.
I understand that a lot of people really enjoy managing teams of agents and are far more productive now. I am happy for them, I really am. Apparently they never actually enjoyed coding, they will show up in droves in these comments to tell you that what they really enjoyed was the end result.
I am more productive too, but I'm going to be honest - I wasn't really looking for more productivity. I am at a point in my life where boosting my productivity is absolutely not one of my personal objectives.
Using Claude Code all day is like being a STEM kid that was gifted a Lego-building-robot. What a miracle! Now you don't have to deal with all the hassle of assembling Legos - you just tell it you want a castle and boom! What a time to be alive!
I think the same will be true of programming and other skills.
I feel incredibly productive. I'm building in an afternoon what would've taken weeks. I spend all my time in either MR's reviewing code, or a terminal prompting.
I've moved from being a coder, to being a product-engineer-architect. I love coding -- LOVE it. But the role has morphed. And, I actually really like the new role. I love being product centric, and caring about architecture -- it suits me wel..
The code Claude produces is often clean, but terribly archictected. So I'm spending a lot of time reviewing in detail.
But the anxienty - OMG the anxiety.
I work on multple features (generally 4-5 work streams) concurrently, so am constantly context switching. I worry about how much time I'm spending on each task. Am I doing enough code review? Too much? Too little?
What about click-through testing? Our test coverage has never been higher - should I be able to just merge these features? Wasn't that always the promise of TDD - merging with confidence? But, it feels like a risk.
And, I'm moving really really fast - but I assume my competitors are too. So that drives me to move faster. Automate more. Use claude in more places. Which brings more anxiety - am I cutting too many corners? Is this all gonna crumble on me?
Also, is everything I'm working on going to become irrelevant? Will customers buy this product, or vibe code it? Do we even have a moat? Anxiety. More anxiety.
My investors are reading stuff constantly too. They're pinging me with market updates, emerging players, shifting landscapes. More anxiety. Also - productivity tools - look - internet influencer X is now producing a gajillion lines of code a second -- are we doing that? Are we token maxing? Are we feature maxing?
Are we building the right number of internal tools? Are we automating enough? Should I evaluate the market for this new tool, or just smash something out with claude? More features. Burn more tokens. Ship ship ship.
Having to explain to investors why we're not going as fast as those guys, and building enterprise products requires care and dedication -- that takes time. Time I should be shipping, coding, featuring, tokenizing. More anxiety.
I'm exhausted.
Anyway, gotta go - herdr has pinged me four times during writing this reply. Gotta ship.
Are people reviewing all the tests that have been created automatically?
What about existing tests? How do we make sure that AI does not fake them (by silent changes or by faking the test output)?
edit: corrected typo
No.
> How do we make sure that AI does not fake them by silently (by silent changes or by faking the test output)?
I have never seen this happen for a couple of months at least.
They're representative of scenarios which your application can enter.
So if the tests were mocked and a bug went live he'd still notice, after all it means your software behaved incorrectly and now you gotta clean up the pieces.
The discussion went from tests being written by ai, to nobody actually reviewing the tests, to the claim that given those circumstances, you wouldn't notice that the tests may not tests anything anymore because of mocks.
My statement was merely pointing out that your previous comment was groundless because they still would've found out about those bugs when they went live.
Maybe. You're assuming visible and obvious bugs.
Absolutely not! Some people ask a code assistant to write only integration/end-to-end tests and commit them. Unittests is only to help LLM generate correct code.
- spending more time upfront working with agent on intent of the change, then design, so agent does not try to work around issues in a way that I don't intend to
- investing in ways in which agent can verify it's own change
- giving access to as much prod telemetry, etc so it can also verify the change post-merge
- obvious stuff, like automated review loops etc
But in the end I still read the "load bearing" parts of the code. Less closely, and I've long ago stopped looking at the tests. It's better for me than couple of months ago.
a recipe for omni-disaster
I spend about 90% of at-computer time in a herdr-like tool.
It feels like playing factorio: whatever is manual needs to be automated away.
LLM writes the test. The test is crap. Update the skill. Ask an LLM to review your feedback on the last 5 PRs and write a skill for a review bot to look for similar problems and post. Now you only see PRs after they've gone through a few rounds of review.
The LLM has stopped picking up new issues. Ask an LLM to diagnose why. Ask an LLM to write a cron to regularly check if the LLM PR pipeline has halted and ping you. Get an LLM to try to diagnose it first.
I'm using a lot of LLMs! Ask an LLM to diagnose why. Write a tool to audit your usage, and produce a weekly report. Have it raise issue to reduce spending.
And like a base building game there are many competing facets, coming all at once.
In factory building games, the systems in play can be understood and are fully deterministic. You're given a set of tools and a problem, and then craft together a solution. Your reward isn't just seeing the item you want on a conveyor belt, but looking at the factory you built ticking like clockwork.
What you're describing is beating out endless nondeterministic fires. It's like you were forced to play Factorio exclusively through an algorithm that designs incomprehensible factories that don't work sometimes, but designs them much faster than you ever could. So you click the magic button and see a mass of buildings appear in front of you. They may work perfectly fine, but if they don't, you need to use the same statistical factory estimator to build an in-game factory fixer mechanism and factory tester mechanism. And then a few updates later the factory fixer breaks so you need to deploy a factory fixer fixer, and a belt clean-up machine, and so on ad infinitum. To me this doesn't seem satisfying in the slightest.
I just focus on one, I use on purpose very short interactions because I want to check every single function directly and not a whole 300lines file or 2000lines pr just at the end. Probably I am going slower than most of you but I feel I retain the knowledge of my codebase almost as good as when I was the one writing the code directly, and my PRs get very few review comments
What get me tired the most is when people submit to me a pr to review and I see immediately they full vibe coded it and just delegate to me the actual hard work of correcting the nonsense written in it with a single prompt without interactions. The worst is when you ask someone why they implemented something that way and they answer with "well calude made it, I really didn't think about it, now that I read it it makes no sense indeed"
* insert melting face emoji here *
Even with short interactions though, the effort in reviewing adds up and I ended up building a tool to help with it: https://github.com/imron/crt
I can review in the tool, and it remembers which files I have seen/approved since a given commit, then I can leave comments against source code and have an agent pick up those comments over mcp so it has context when making corrections.
Sometimes I feel please someone give me a job where I can work 12 hours a day - but the rest of the 12 hours I don't need to think about what will happen tomorrow.. It will still be less stressful.
Why do you dream about working 12h a day when 8h is pretty standard.
I ask it for a quick complexity thermometer reading when ever the rabbit hole is looking like it might explode.
Also walking away from it every few hours and doing random simple mindless stuff has helped. Like I will go inspect my fish and plants. Try to have a chat with the cat, get others around to do a little dance etc etc.
I have a priest in the family who has to deal with lot of heavy duty stuff/people's problems. They train them not to get totally sucked in and loose themselves. He calls it grounding. Just simple mundane things to focus shift out of anything heavy.
For me, the conclusion was: I want a spatial layout of work on a canvas per project and narrated demo videos, like in a feature launch review, where I as "engineering director" can decide to launch or give feedback for another iteration. Input mostly by voice. So it's a bit like sitting in a meeting room with engineers coming and going.
As software is so cheap to build nowadays, I think you just need to find out how you want to work and then build the harness for that style of work.
(If you want to try out my style, it's free and open-source. I'm sharing it because I'm excited about how much fun it is to use: https://github.com/sadilek/obeya)
FWIW, there might be a silver lining in this. A lot of dysfunction in the current state of tech can (IMO) be traced back to violation of boundaries. This might lead to a cultural shift. Though, that might also just be wishful thinking on my part.
___
> I would love to hear your tips, but still be relevant in this AI era, so my boss does not fire me.
If your boss fires you because you didn't multiply your performance by doing more work in the same time (without your compensation changing), that's not your fault but simply your employer trying to squeeze you harder and that not working out.
Always keep that in mind in these discussions.
Which isn't to say that you should not improve efficiency and all, but it needs to be organic. If something inorganic is asked, then the (moral) failure is on the asker and not the askee.
Now instead I don't get to code anymore and I'm the manager of a half-dozen mid-level engineers with unflappable confidence and very, very poor long term memory.
What helps me from time to time is "mental offloading" - I basically sit for an hour and write down everything that bothers me, all unfinished thinkings and questions. And I'm not trying to solve them, I'm just putting them to the paper (or on my laptop / phone) to offload my mind and later I think about them in a more detailed way. I hope that helps!
AI-generated code works great for me, most of the time. I'm productive, I ship, I can show stakeholders stuff in no time and and they are mostly happy.
But the development sucks. I don't even really look at the code anymore, because reading generated code is a drag. And there is no reason really to do it. There is no fulfillment of "I solved this real problem using some intricate logic that came out of my mind". I just throw tokens at it, and it works, great.
I don't really talk to coworkers anymore about the code. We don't do coding sessions together, or bang our heads together over some hard problem, because there are no hard problems. We don't learn from each other anymore. Everyone just asks their agents, and they are fine.
I feel my brain atrophying. I forget how to do things. Developing now feels like being kinda drunk.
Can this be solved? Absolutely: just do it like you used do! Just like quitting to smoke is easy, just don't do it, right? But what if all your colleagues go out to smoke all the time? Do you sit alone? Do you go with them, but just stand next to them in their fume? Or do you need to find new friends? Going back feels like going back to an Abacus after you invented the pocket calculator.
The AI era is, so far, not a problem. Obviously, my having retired is significant.
Last night I spent an hour with Claude (free tier) setting up a "modern" 6502 build environment for MacOS. By the time I went to bed I had 1) 6502 assembly syntax highlighting in VSCode, 2) was compiling 6502 assembly for the KIM-1 to generate a "paper tape" format, 3) pasting this paper tape binary into CoolTerm to send over TTY (1200-N-8-1) to my KIM-1 clone (PAL-2).
(If I'm tired it's because I excitedly stayed up later than usual to get Claude to help get this workflow working.)
On the other side of the world… a friend who is still at Apple tells me how other engineers at Apple are using Claude to make changes, create pull-requests on code he maintains with a, "Hey, can you approve this change?"
Now he has to go through the code diff that Claude generated…
Oh!, there's some stupid shit in that diff. Did the engineer who submitted the pull request even look at the code before submitting? Who knows. Either they did not or they were not qualified (or well-versed enough with the code) to spot the AI-fluff/crap.
So, yeah, that would suck—keep me busier than I would have been Antediluvian (thanks Claude, for suggesting that made-up Latin).
Of course not :D
Claude Web works differently than Claude Code. I use Claude Web/Chat for briefs, general overview of the project outside of code/app, for planning and handing off the work to Code, that includes prompts. Code is tasked to output a detailed view of what it found, run measurements and tests (designed both by me and Chat) and report findings back. That is fed back to Chat and I get a rough summary. Errors are categorized by their blast-radius and so on.
Eventually I do a deep dive of what changed, how well it's handling, how much it's struggling, token usage, time spent per review/test and so on. (btw, Code is tasked to run an independent review via contextless agent on the codebase, that easily costs me another 20-30 mins and ~200K tokens on my current project, each run).
It still makes architectural issues, tries bandaiding and patching instead of properly designing, overvaluing/devaluing certain cases and misallocating where most of the focus should be, but that's relatively easier to spot and propose a fix for than doing a line by line code review.
Codebase quality is ensured through mapping the flow and the output space and having a hard-pinned tests Code can't modify to green its answers, + the independent reviewer I mentioned above.
It's not as fun as going in by hand on deep technical problems, sure. But AI usage is less about fun and more about output speed, so use it for what it is for ig? :D
1. Being okay with being "less" productive. I'm probably still 10x more productive than I was in the pre-AI era, but in the early days of AI I would have eight Claude sessions running. Now it's two or three and I take more advantage of the free time I have between Claude assignments. My bosses are still happy with my output and I'm not being burnt out.
2. Very personal, but building my own tooling really helped. I built https://smalldocs.org / https://github.com/espressoplease/smalldocs (a document creator for AI agents) which allows me to take a lot of work outside of a chat history. E.g. I might make 10 MRs and then say "smalldoc me a list of the MRs" and I have that in my library, or if I'm considering various architectures I'll ask "smalldoc me a simple overview of the three architectures with mermaid diagrams" and then I can review the agent's thinking in a cleaner way.
But it’s not because I don’t think it’s important and that I don’t want to do it. I have a very engineering mind that likes to be thorough, and I feel like with AI I get to do all the “boring” parts, without it being boring. I get to spend more time on the architectural thinking and finding tests for edge cases, focusing on stakeholder needs and keeping the momentum by developing fast.
The biggest problem is that I’m losing grasps of the projects. I don’t review much anymore and I feel like I’m losing my ability to code.
I realised how easy it is for me to build the dash.
But how hard it is to sit down for 1 hour, look at the pages, scroll through them as if I am my client, notice inconsistencies, notice where the numbers don't make sense.
Sitting down, and actually checking and comparing and thinking were already hard before AI, and they became a lot harder.
I used to think that only the workers (the coders, burger makers) did all the hard work, while the managers and up basically sat on their butts and relaxed. But decisions and managing even without coding is just as exhausting.
Now I'm down three quarters of my traffic and my inbox is filled with LLM-generated outreach. I'm competing with AI trained on my work, and with slop websites generated with that AI.
The original research is still necessary, but it has to happen with a dwindling audience and a shrinking revenue. I have to worry a lot more about paying rent, so I have less time to contribute to society.
Worst of all, LLM noise has filled every crevice of my life. It's billboards, blogs, music, videos, social media. It has filled the world with spam and nobody is happy, but everyone's Christmas bonus depends on adding oil onto that fire.
I mourn the loss of our internet, our culture, my career and my business all at once. And for what? We're all worse off!
I think part of it is just a point in time thing though (hopefully). We can’t trust the agent output so need to check it. We can’t trust the agent to use common sense so have to think through every clarification instead of assuming a competent peer will come back with any necessary questions instead of making bad assumptions. I think all of these things will continue to improve over time. For myself AI quickly went from so painful to being avoided to something I can’t work without.
Yes, reviewing an LLM code is at times infuriating, but at least I can be brief and direct, rather than spend 2x the effort trying to make sure it won't be interpreted as harsh.
I went back to working on one project at a time to regain my sanity.
What was making me tired was all the doom and gloom online. I stopped consuming certain news and stopped going on Reddit entirely and my outlook is noticeably better.
I haven’t cut out sad news entirely but I don’t read speculative opinions about things that don’t actually affect me.
In terms of the overload I find that the clearly defined problems have much more obvious solutions. Identify your pain points and use AI to build you a tool or reusable prompt to make your life easier. It’s a constant refinement process to figure out what helps you.
For example, I was having a hard time managing multiple sessions so I had it build a window manager with notifications and color coding based on session hooks and git issue/PR status. I’m comfortably up to 8-12 sessions working in different worktrees.
Otherwise I've never been less stressed and chill in my working life.
I'm not sure if you could share the window manager but I'd love to get a glimpse of it, as it sounds interesting!
- accept that you cannot be concentrated for hours and that reading an AI chat is mentally challenging (see first point about breaks)
- Talk with your coworkers about it. You are not alone, it is important to talk about how you feel
- set some boundarys, e.g. I don't work in parallel on different stuff
- do NOT doom scroll while the AI is "thinking". Try some meditation or mindfullnes exercises. Doom scrolling on reddit, HN or youtube will mess up your brain even more.
- from every task, try to do something on your own, and then scale this up by AI. That tip is from a former co-worker. Their idea was that you need to figure out things first and then let AI do the boring task to scale it up. Depends on the topic sadly.
- do some things by hand, especially if they are fun for you. E.g. I really like UI design, somethign that basically is gone now for me.
It feels very strange on multiple levels.
There is the tiredness of reading text all day long. The fear of loosing my coding abilities because Claude does is. I am totally loosing track of what I did when. I can tell you about things I coded 5 years ago, I cannot tell you what I did with Claude last week.
And don't get me wrong, I LOVE planning stuff out before doing that. If there is anything positive about AI, then it is people actually PLANNING before doing.
On the other hand, I can now do/try out things I did not do because I did not have time. But at the same time, it feels like I am working on making myself not needed anymore in the future.
Worst thing, I feel like I am getting dumber. My attention spans or the "I need to know why this happens" span gets shorter, because with AI everything has to be faster?! Riiight?
I don't know how this will turn out in the future.
I see people underselling themselves becoming anxious about working without AI (either because of how they think about their abilities or management pressure).
I am getting chat snippits from AI conversations via Teams, PR checks are skills now and we need 3 rounds of discussing findings with AI ...
We live in interesting times. But yeah, I am feeling with miserable about it.
I used to grind so hard, but now it's so chill. I don't mind this change.
The main thing i dislike is how much my profession has been devalued. Code is worth so little now. It's sad.
Its crazy to see how we can now do things which were not thinkable a few years back.
Finally i read real progress i can see and use by myself and not constantly some crypto and blockchain stuff.
While I would like to have a stable world/live and for sure AI is disruptive as hell to a lot of people and potentially to me too, at least progress of technology i can't stop.
And i thought about automatistaion for decades for sure, I don't play devils advocate and thinking "yeah its great that we invented the loom" while hating AI for what it is: tech.
In worst case I only watch one 45 minute ai summary video from my YT channels, but I do have a lot of random small experiments i'm playing around with anyway.
What i discoved though: if you never go above 20$ subscriptions, you hold yourself back because you don't think about approaching bigger experiements.
Also, trying to put more and more focus time, in which I don't allow myself to use AI.
I find it very rewarding.
For me it's very exciting and AI has been so far a crazy enabler for me. Both in my work and my personal life.
But I also enjoy keeping up with it and reading about it etc.
Claude's default personality is a bit too combative for me. Another instruction I've given it is "don't argue to win".
LLMs appear to be able help me get through that. I'll take the help.
BUT JEEZ! Dealing with "Rain Man"[1] LLMs week-in and week-out is wildly annoying, frustrating, and anger inducing.
If they all would just STOP sounding so authoritative when spewing nonsense!
[1] After the Dustin Hoffman character in the movie of the same name. "Ah but you know I'm an excellent d̴r̴i̴v̴e̴r̴ programmer."
That your workplace, not "AI era".
that 'dead time' was when my mind was doing background processing, so when i reached another decision point, i was mentally prepared. It wasn't actually dead.
and now, working with agents, i feel much much more tired because that background processing cant happen
This is something I avoid at all costs. If it's someone else's slop I just won't read it, and if it's "my" slop I prefer to interactively ask questions about it rather than reading a wall of text.
It didn't really bother me during the altcoin/nft scam era, other than from time to time trying to explain at a normie that they were mostly scams, you know the concept that if you want to sell an asset for X you have to find somebody willing to buy.
But at least you could simply ignore all the mentally ill megalomaniacs that were like duude this is the future
This time instead it's so pervasive that you can simply not excape the narrative and simply concentrate on the good stuff about the technology. It's the narrative, the pressure, the whole global situation, not AI in general
In true drinking the kool-aid style, I've attempted to solve this problem with an LLM!
I have a skill which aggregates what I've been working on over the last day (from slack, calendar, mail etc. MCP) and writes a notion journal entry for me. It runs in the evening when I've finished and is ready the next morning when I start. It helps with the context switching - reminding me of deadlines or what I was in the middle of before I left.
The actual using of AI, though, doesn’t tire me much. But I mostly use it for answering specific types of questions that I’d otherwise need to spend a lot of time researching.
More of what? I see many here using AI to build more of the same things in another language such as rewrites and then claiming that they are "productive", which is the problem of many engineers here.
Not that much risk taking if all what everyone is doing is just full blown rewrites and then abandoning it in a couple of months later and just wasting tokens for free.
If the result is to vibe code something for free and it makes no money and has no users then it is a waste of time.
Clearly there are few limits to what you can implement with AI agents today.
Depending on the complexity, you will spend a lot of time QA testing, giving feedback, and iterating.
One can take shortcuts such as porting existing software, as you mentioned. Many are just content publishing crappy novel software that they barely tested.
You don't need to do that. No one is forcing you not to QA test and iterate until you have a good product.
rest i use to think
But then I have to repeat this lecture to so many people it’s wearing me down. I’ve mostly resorted to just not reading it and simply ignoring any AI text sent to me.
That't the key. Not AI.
I would recommend you to try using FIM/line completion rather than the agent approach. Using FIM it makes me (feel) productive while still being in control. Oftentimes when writing code for production, documentation etc. the agent approach feels like a wash to me - sometimes it saves time, sometimes it wastes time. With FIM the gain/quality is more consistent.
It's completely destroying genuine conversations about the pros and cons of AI when people will just forever parrot that everything is slop
Yeah. Quitting before this shit eats me alive. I'm in an EE degree program now. Better three years of study while everyone else gets to deal with the bananaware and the bullshit, and let's see then once the seas have calmed.
Until then, I actively try to avoid engaging with Slop Generator regurgitations and call it out whenever I see someone not disclosing they used a Slop Generator.
AI is a boon for learning, but the result will heavily depend on mental capability of the user; having a depression-induced brainrot doesn't help though.
And, in the last year clankers have started coming for us EEs.
The reality is that many people are more fucked than I am, but I can't stop the bad thoughts.
Get your practical skills in residential and commercial electricity up to speed and join the trades. The world is going for solar, heat pumps and electric vehicles no matter what corrupt politicians want (yay free market for once, lol), and that means there will be demand for people actually pulling wires and installing cabinets for decades to come. Start out as a side gig and when you find the clankers get too close to home turf, expand.
It also helps that a lot of my AI use is on my own side projects, where I set the pace. That makes it fun instead of draining. Maybe the tiredness has less to do with AI itself and more to do with being pushed to maximize throughput?