Today's goal, try not to micro-manage the AI coding assistants.

I'm pretty decent at walking a coding agent through to completing a PR for an issue. I can identify documentation needs and guidelines. I get work done fast. But that's table-stakes these days.

Great article on expertise and the challenges of a novice: boydkane.com/essays/experts

Going to live... sky? What do we say instead of tweet? Anyway, going to live-post as I work through some llm blueprint recognition stuff.

It's nice that llms don't get frustrated. Some of the things problems they work though I would not ask a human to work through.

Got into the "instruct your LLM to call you x" thing to track when rules are and aren't being used. Currently I'm being called 'dude' which has been nice!

I used to only do data. Never once got asked for a UI. Now I do data and GenAI and somehow endpoints/data isn't enough. My GenAI is just more data, folks! I'm not a good choice for UIs. I have that humility, but they have such faith :-( So I develop a crap UI and they're like "Great!"-- It's not!

Had a solid programming session today. Set one cursor instance off on its own for a good twenty minutes or so working on a refactor from something lightweight into something more sturdy. Other than continue-clicks and a single course adjustment it got the job done with my focus on another project.

I realize I have two modes when coding with AI
1) Knowledgeable growth where I have a good feel for the code
2) Unrestrained prompting the agent until it just works

2 rarely works out quick enough to justify not taking approach 1, and even when it is quick enough I gain nothing for my toolkit.

I officially dislike snowflake's llm interface. Their API is lacking. Their documentation is lacking. I feel like it's a year ago and they're just starting to fiddle with this.

I wish more things worked with my horizontal mouse wheel.

Still working out kinks in my AI-Coding workflow. Trying:

meta-process-template -> process-template -> process

Where process generates the overview, specs, division-and-prompts-for-specs and tracks the progress thereof via the process-template.

Will update after a few more runs.

Completely refactored my AI Assistant's data layer and file layer. My old code was terrible to behold and is now replaced with tidy and functional code. Looking forward to being appalled by this code in six months!

Minimum-viable-products are like I'm making a company's tech tree. You sink the time to get it up and then have backlog items like:
Enhance customization - 1 story points
Improve accuracy - 2 story points
UI Integration - 5 story points*

*Needed to unlock ShinyMarketing Level 2

AI PoCs can be flashy because they ignore the details. Summarize a document, sure. Categorize the document and apply summaries appropriate to the portion? Have consistent categorization? Complexity is tripled.

Man. Using Cursor dramatically shifts the cost-to-effort ratio for unit tests. They're actually fun and interesting instead of needed-but-dull and I can iterate damn fast.

Caffeine and almost-too-spicy food is a great rush for coding.