Notes
Short reads, opinions, things in progress.
The half-finished thinking — written as it happens, not after it's tidy.
Rules first, AI for the residual
On Podsque's completeness pipeline, deterministic rules filled ~70% of the gaps at zero cost and the model only handled the ~30% that needed judgment. Reach for AI on the part that's genuinely ambiguous, not the part you could just write down.
Make the AI show its work
Document X-Ray doesn't just extract a field — it returns a confidence tier and the evidence text it pulled from. A score or an answer you can't interrogate is one you can't trust or defend.
The spec is the review surface
Once a model writes the code, the only thing left for a human to meaningfully review upstream is the spec. So write it like it matters.
A prototype beats a wireframe for validating requirements
A wireframe shows a screen but can't prove branching logic, so 50 functional requirements stayed theoretical until product, QA and dev could click through a working flow. Validate the behavior, not the picture.
The spec is a living document, not a deliverable
The Podsque catalog spec grew to 41KB — 50+ field schemas, 10 locked business rules — because it stayed the place decisions got recorded, not a doc filed once and forgotten. A spec you've stopped editing is a spec that's already wrong.
Stop asking which model is best
It's the wrong question. The useful one is which model does which part of the job — and how you hand work between them.
Notion is a perfectly good backend
Half my pipelines write to Notion — not because it's the 'right' database, but because the team already lives there and can see the data the moment it lands. Pick the backend your users already open, not the one that benchmarks well.
Ship it on a schedule and forget it
The barcode enricher runs every five minutes on launchd; meeting tasks sync every thirty. The unglamorous win of AI tooling isn't the model — it's the cron job that runs it with zero manual intervention.
Half-finished, on purpose
A note on why some of this is shared before it's polished. Working in the open means showing the thinking, not just the conclusion.
Prompts rot
The prompt that worked last month quietly stops working as models update. Treating prompts as artifacts means noticing when they go stale.