B ack in January, with the frontier models ramping up their coding capabilities, I got the itch to sit in front of a computer and start building something again. My goals in ORDER BY priority DESC were:

1. Learn how to work with AI.
2. Learn Rust and make it my primary programming language.
3. Have a single project I can use as a test bed for whatever research path I wish to explore.
4. Build something semi-useful for myself and close friends & family.

I settled for building an AI assistant, not because the world needs more AI assistants, but because it is a well-understood application for AI and a wonderful test-bed to experiment with whatever technology I wanted.

My journey with AI coding mimics that of a lot of developers, starting with copy/pasting from chat interfaces, Co-Pilot autocomplete, agentic development on a function-by-function basis and as of summer 2026, settling into the following worflow:

1. Roughly spec out the feature I want to implement and put some bullets in Obsidian.
2. Point agent, usually but not always Claude, to spec and ask for a plan. Ask to review plan with a model from a different lab, i.e. Codex.
3. Read plan carefully.
4. Ask corrections to the plan.
5. Ask for implementation.
6. AI & Human code review (split 80/20)

I have the luxury of not having pressing deadlines and only working when inspiration strikes, and more often than not, I get stuck on step #3. Not only stuck, but it drains the living soul out of me. It takes 3-4 days to get through what should have been 3-4 hours of work or I just give up and move to implementation without a thorough plan review, which usually results in messy code.

I’m sure this will be a non-issue in a year or two and this blog post would be outdated. But as it stands, reading LLM-isms is an exercise in frustration. Nobody talks like that! So I changed my workflow:

1. Involve AI early and together we spec out the feature I want to implement. I write the spec in Obsidian in more detail and in my own words.
2. Ask for a plan.
3. Skim through the plan and usually move to the implementation right away.
4. Go and read the code in my IDE, rewrite all comments and potentially refactor code to make it “tighter”.
5. AI Code Review

Spending some more time in the spec stage, fleshing it out more with AI help but making sure I write it out in my own words means I can just skim the plan for any glaring omissions or large deviations from the spec. And in step #4, using code instead of English as the common language between me & the AI is a lot more pleasurable and anecdotally results in better outcomes.

And the seam? Well, that’s an honest take that doesn’t earn its keep.