full-stack fatigue in the LLM era — the part nobody's saying out loud

i've been writing code for money for a long time. i've never been more "productive" than i am right now. and i've never been more tired. ngl, something about this era is cooking us and nobody wants to say it out loud because the line goes "AI is amazing, you should be shipping 10x more" and if you push back you sound like a luddite who's about to get replaced.

so let's say it out loud. full-stack dev in 2026 is exhausting in a way it wasn't in 2022, and it's not because the work is harder. it's because the work changed shape.

1. the "thinking time" got stolen

old loop:

• read a ticket
• stare at the ceiling for 20 minutes
• have an idea
• write it
• it doesn't work
• think more
• fix it
• ship

new loop:

• read a ticket
• paste it into Claude / Cursor / Copilot
• get 80% of the answer in 4 seconds
• spend the next 2 hours figuring out why the 20% is subtly cursed
• realize the model made up a function signature
• fix it
• ship
• next ticket. immediately.

the staring-at-the-ceiling part is gone. that part wasn't slacking — that's where the actual engineering happened. that's where you realized "oh wait, we shouldn't even build this, we should refactor X first." LLMs collapsed the time-to-first-draft so hard that the time-to-think-about-the-problem evaporated with it. you're just always in execution mode now. it lowkey wrecks your brain.

2. you're a code reviewer now. forever.

full-stack dev used to mean: write frontend, write backend, write SQL, write deploy config, get good at all of them over years.

now it means: read AI-generated frontend, read AI-generated backend, read AI-generated SQL, read AI-generated terraform, catch the four bugs the model planted in each, integrate them, ship.

reviewing code is more tiring than writing it. there's a reason senior eng burnout was always a code-review problem. and now your job is basically perpetual PR review on code your "junior" (the LLM) wrote with confident energy and zero context. no cap, my eyes hurt different now than they did three years ago.

3. the stack got 3x wider in two years

i did a count last week. things i'm "expected to know" in 2026 that didn't really exist or matter in 2022:

• vector DBs (Pinecone? Weaviate? pgvector? Turbopuffer? Qdrant?)
• LLM orchestration (LangChain — sorry, dead — LlamaIndex, raw SDKs, Vercel AI SDK)
• prompt engineering (sigh)
• evals (which are basically tests but vibes-based)
• RAG pipelines
• agentic frameworks (every week there's a new one, they all break)
• MCP servers
• streaming SSE infra for AI responses
• token cost monitoring
• model routing / fallback chains
• the entire universe of "AI gateways"

this is on top of: the frontend stack that already had React + Next + Tailwind + 6 state libs, the backend stack that already had Node + Python + Go + Rust depending on the day, the cloud stack, the observability stack, the auth stack. the "full" in full-stack just got wider. nobody's talking about it because admitting it makes you sound slow.

4. the goalposts moved. they don't tell you they moved.

what "a feature" used to mean: a button that does a thing, validated, tested, shipped, two weeks.

what "a feature" means now: a button that does a thing, plus an AI-powered version of the thing, plus the AI feature should be streaming, plus it should have a fallback model, plus the cost should be monitored, plus the eval harness, plus it ships next sprint because "you've got Cursor."

every PM in the world saw the GPT-4 demo in 2023 and silently 3x'd their expectations of what one engineer can do in a sprint. they did not 3x the deadline. they did not 3x the comp. they 3x'd the scope and called it "leverage."

5. the dopamine got weird

writing code used to feel good. you'd struggle with a problem, crack it, get the hit. that hit is what kept you going through the hard parts.

now: prompt → output → it works → "huh, ok, next." no struggle, no crack, no hit. it's like eating a candy bar instead of cooking a meal. the calories are there. the satisfaction isn't. you finish a sprint that would've taken a month in 2021 and feel… nothing. lowkey concerning. i don't think i'm alone in this.

6. the imposter syndrome got a software update

the old imposter syndrome was "do i actually know this stuff or am i faking it." you'd push through it because eventually you'd solve something hard and remember oh yeah, i can do this.

the new imposter syndrome is "do i know this stuff or is the model knowing it for me." and there's no clear answer. did i write that regex or did Claude? did i debug that race condition or did i just paste the stack trace? am i a senior engineer or am i a really fast typist with good taste in autocompletes?

it doesn't matter that the honest answer is "you're directing the work, you caught the bugs, you made the calls — that's still engineering." the feeling persists. tbh i think the industry's collective mental health is going to be a problem we look back on in five years.

7. so what

i'm not saying turn off the LLMs. that ship sailed. i use them daily and i'd be slower without them. but a few things i've started doing that genuinely help:

• protect the staring-at-the-ceiling time. i don't open Cursor for the first 20 minutes on a hard problem. i read the code, the ticket, and the related stuff. then i pull in the AI. the thinking has to happen somewhere.
• write the spec, not just the code. if i can describe what i want in english tightly enough that the model gets it on the first shot, i actually understood the problem. if the model keeps missing, i don't — and that's a signal to slow down, not to re-prompt.
• rest. like, actual rest. the new dev loop is denser. you cram more cognitive load into the same hour. the recovery time has to scale with that or you crater. 4-day weeks aren't a slacker thing in this era — they're a "the work is more concentrated" thing.
• build something dumb for fun. with no AI. by hand. it's the only way i remember why i started doing this. you don't have to ship it. you don't have to post it. just feel the old hit again sometimes.
• say it out loud. the worst part of this exhaustion is the gaslighting — the "you should be shipping more, this is the golden age" energy. it's ok to be tired. the work changed. acknowledging that is not a skill issue, it's pattern recognition.

we'll figure out a sustainable shape for this eventually. right now we're in the bit where the tools sprinted ahead of the practices and the humans are paying for the lag. you're not crazy. you're not slow. you're just early to noticing.

(if you read this whole thing and felt seen — go close your laptop for an hour. seriously.)