Even though I’ve been coding for over 4 years now I still fell for the marketing hype behind AI and coding agents hype.
"If you’re not using AI you’re going to fall behind"
“AI is going to replace you in 6 months”
After trying it on and off for years I decided to make my own conclusions from my experience rather than listening to the AI CEOs constantly who are just marketing their products.
The unlock being that even though the hype cycle keeps saying the agents can do things autonomously, the reality is that if you Treat the coding agents as juniors, you become so much more productive. It’s just a better autocomplete.
Side benefit of AI forcing you to simplify is that you’ll be a better engineer for it.
The Reality of AI in Mature Codebases
I don’t know what magical claude.md files people are writing but I have had 0 success with being general with my agents. It desperately needs guidance and direction.
How I use Claude Code and Cursor Agents now
-
Have an end goal in mind for myself
-
Start in plan mode and break down the task into smaller parts
Be thorough in the plan. Think in terms of product and outcomes
-
Specify Exact files
Once there is a general plan, go through and make sure the agent knows exactly which files it’s going to change and what changes it’s going to make.
-
Get exact code changes
A lot of the time it won’t actually give snippets so this has to be iterated on. I prefer to get this done in the plan phase because once it starts building, its very difficult to readjust
-
Don’t be afraid to throw away all changes and start again
Its quite common for the agent to not get things right and either go off course or the implementation is not to your taste. Until there is a better version control to work with AI, use git and just throw away changes
Why This Makes You a Better Engineer
Forces Smaller Commits
I’ve switched between Claude code and cursor and opencode. Cursor is the best with seeing the changes made and the diffs so its my preferred choice right now. Still I end up using the git diffs to see all of the changes and then commit them before starting on the next small task.
Sometimes I like part of the implementation so I’ll only stage and commit those parts. Then re-try the rest of the implementation on a fresh commit with a fresh agent.
Improves System Thinking
This really lets you take a step back and think of the whole system. Its really helped me architect my code a bit better because its so much cheaper to write code and refactor now
Develops Mentorship Skills
I’ve never had the opportunity to mentor juniors, yet, so I’m hoping I can pick up some of the required skills like breaking down tasks into simple steps.
Real Example: Feature Flag Implementation
I had a task to add a feature flag.
My Input:
- Exact file location
- Variable name I wanted
- Specific implementation approach
- Context about the existing code structure
While Claude implemented the code, I went to PostHog, to create the feature flag and variants. By the time I came back, it had implemented the changes as well as using our feature flag naming conventions
Total Time: About 30 minutes from start to test and starting deployment.
Final Thoughts
AI isn’t coming to replace you any time soon but I do think there is value and leverage in being able to use AI effectively to work faster. This is coming from someone who loves the craft of software development.
I’m just trying to reframe it so that I’m not just using AI but practicing all the skills needed to improve and hone my craft.