The Workflow Around AI Matters More Than the Model
In my recent posts I shared how I plan and build features with AI. Now I want to talk about how I use it beyond generating code.
I'm using skills for pretty much everything. I basically don't do anything manually anymore, except testing. And even there, AI helps a lot.
AI handles the end-to-end of work now. Some examples:
Analyzing GitHub issues
- Reading issues from a specific milestone
- Checking and updating issue progress
- Creating sub-issues when needed
- Adding follow-up comments
- Checking all pending items assigned to me
Quality review
- Complete review of different aspects of the code
- Checking good practices and respect for codebase patterns
- Security breaches
- Performance gaps
- Compatibility with Meteor 3
- Specific project rules, for example: all code from the shared folder should not depend on code outside it
It's important to note that just having markdown files is not good enough. AI needs tools. For quality review, for example, the AI runs lint, unit tests, bash scripts, and more.
Again, AI will not replace devs. It still needs to be guided. Same way a car doesn't replace me, it helps me reach my goals faster.
The main point is what's around AI: the good practices, the workflow, the harness, the environment, and the project rules. AI models alone are not good enough, even the powerful ones.
Before you blame AI, create a good setup for it. Agents can't read your mind. Facilitate their job. Tell them what you want, how you want it, what you don't want, and how to measure success.
See you next time!