For anyone who knows of me, they know that I don't like using coding agents. I have nothing particularly against them, I just don't prefer them. I like the quality and control of workflows, in a direct chat window. You can see as much in my old (and now out of date) guide on how I develop with LLMs.
Each of the below PRs represents a weekend's work. I start Saturday morning, and I go until Sunday late night. Of course I take breaks to run errands, eat food, goof around or watch movies, go places, etc. But every other bit of free time in my weekend is coding like a madman on Wilmer.
- This weekend: https://github.com/SomeOddCodeGuy/WilmerAI/pull/76/files
- Last weekend: https://github.com/SomeOddCodeGuy/WilmerAI/pull/75/files
- The weekend before that: https://github.com/SomeOddCodeGuy/WilmerAI/pull/74/files
- The weekend before that: https://github.com/SomeOddCodeGuy/WilmerAI/pull/72/files
None of this is from a weekday; I just don't have the time or energy after work, since I generally work late. I usually just study on the weekdays, test out new models, test Wilmer some and play around with workflows, unwind a bit, etc in whatever time I have after work. I'll keep a todo list through the week of what I want to work on when the weekend comes, and then on Saturday morning I hit the ground running.
Agents are fast... but don't discount chat-window coding. I can move this quickly because I know the codebase like the back of my hand, and that comes from having held the LLM's hand every step of the way. The LLM gets input on the architecture, and you'll definitely see that in some places, but I put eyes on almost all of it.
There are some parts of the code I only skimmed- it's true. I mean, my whole tech persona online is devoted to building and using AI tooling for development... y'all should have probably figured I use AI for development =D And when you're code reviewing that much code in that short amount of time? Of course you're going to end up cutting corners.
But for the overall architecture? What each script and each component and each feature does? If I want to add a new feature, I can tell you off the top of my head exactly what files need to be touched to do it, and what parts of the files need to be updated.
On a heavy weekend, I'll generate over 100 individual conversations (I love single turn convos for coding; I hate going into multi-turn if I can avoid it. I'll send a question, delete the convo, start a new one).
So yea... not sure if this is 'vibe coding' or not, but give me a good workflow or proprietary LLM in a chat window and I can tear through some code pretty quickly. I'll take that over an agent most days... most. Some things its nice to pawn off, though lol.
Top comments (0)