I’m a SWE with ~5.5 years of professional experience now and anecdotally see AI used primarily by juniors who use it as a crutch. Moreover, the vast majority of the “best” engineers I know do not use any AI-assisted coding tools (e.g. Copilot). They do, however, occasionally use LLMs as a search engine for unqualified questions (I.e. where they identify that there are unknown unknowns). Is my anecdotal experience representative of reality? If not, I’d love to hear peoples’ workflows (especially the historical high performers)
11 comments
25 years experience. I don't use it day to day at my work because not using it is easier. Don't use co-pilot but did try for a period and felt it got in the way. I use llms as a search engine at times. For personal projects I'll use it for content generation, deep knowledge dive downs, linux scripts... I also use it to generate code but if I know the look/ shape I want then its a matter of teaching the ai if I don't know what shape I want it's a journey.
I’m a staff-level FE for 8 yrs. My workflow since 2024:
1. Exploration: LLM first, docs second—cuts discovery time by ~3×.
2. Boilerplate: AI generates, I refactor on the spot; never merged blindly.
3. CR: bot leaves a first-pass checklist, humans focus on architecture.
4. Legacy spelunking: 200k-context summary + mermaid call-graph.
5. Rule of three: AI writes glue, I write core, tests cover both.
Result: 30-40% more features shipped per quarter without quality drop.
I'm 50% sure this was written by an LLM.
Use it to get a primer in a new area If i'm debugging i'll feed it what i know and see what it gives me. If nothing else gives me some ideas to get started. Generate test data
I've tried it to generate html/css for an email and it kind of works but depsite asking it to doesn't work across all versions of outlook and gmail.
I'm overly cautious about what I paste in. Just like how you can find PII data in logs I think the amount of PII data that's being pasted into AIs will be crazy.
I’ve been a senior SWE (~10 years in backend and infra) and I’d say your observation is partially true. Juniors often use AI as a “crutch” — they rely on it to fill gaps in fundamentals or recall syntax. Seniors, on the other hand, use it strategically:
Code review & refactoring assistant: I use AI to sanity-check my design or spot potential edge cases.
Exploration & learning: When evaluating a new library or framework, I ask AI for comparisons or best practices.
Docs summarization: LLMs help me parse long RFCs or documentation quickly.
Prototyping / boilerplate: For scaffolding boring repetitive code.
But not for actual algorithmic thinking or critical code — those still rely on human judgment.
In short: top engineers do use AI, but they use it like they use Stack Overflow — a tool for leverage, not a crutch.
I find it really usefull for simple but time consuming task (mapping of fields for instance) or for precise questions.
I see some people using it to vide-code whole features which often lead to code that work on the surface but when you deep dive in what it actually does it's catastrophic
ITT: "Juniors" thinking they are "Seniors" (which by itself are almost useless terms)
I agree the label often means different things to different people, but ultimately I’m curious to determine if my anecdotal experience of AI being used very sparingly by the top 10% of performers is characteristic of the behavior more broadly, or just an anomaly in my circles. Do you have any thoughts on this?
I use it as a pair programmer some of the time, especially in areas that I'm not super knowledgeable about, like arcane configuration details. I just use the ChatGPT app with cut and paste; I have not yet graduated to AI IDE tools. I'm thinking about it though.
Its great to get back into programming when you have become rusty because of managerial duties.
I am a lead software engineer with 12 years of experience and a recent master's degree in cybersecurity.
I use GitHub CoPilot everyday. I usually limit it to a fancy auto-complete. It's really helpful for repetitive refactor tasks, I just have to go to each line and it updates. This way I still see what's happening, but it is faster than manually making those changes that can't be solved with a find and replace.
Sometimes I fight CoPilot because it will continue to suggest something I do not want to do. In those instances I code faster as to outpace the AI's latency.
Other devs are building MCP servers to help access our tools for AI integration. Devops seems heavy on spec driven and test driven development through AI. The spec driven development looks interesting, but a bit of overhead to get started.
Our company has an AI first directive right now where we're supposed to use AI for everything and see what works. I somewhat disdain it, but it's also fun to have a directive to try new things indiscriminately (using AI). The more I drink the Kool aid, the better it tastes.