AI
Small models, big prompts: the operator's edge.
Most agent demos use the largest available model. In production the right move is usually the opposite — and a much sharper prompt.
06 — promptsThe setup
Most ops teams don't need more software — they need the right pieces wired correctly. On a recent engagement we inherited a stack with eleven different SaaS tools doing roughly the same job. Cleanup wasn't a feature, it was the work.
Why it matters
When integrations fail silently, every other system inherits the bug. The fix is not "add monitoring" — the fix is to make the integration single-source-of-truth so there's no drift to monitor.
Build for the system you'll have in six months, not the one you have today.
How we did it
- Audit the moving parts. List every entry point and every webhook.
- Pick a backbone — Postgres, Supabase, or just a JSON contract you actually own.
- Migrate one feature at a time. Don't rewrite, replace.
// Single source of truth — the rest of the system reads from here.
type OrderEvent = {
id: string;
status: "received" | "fulfilled" | "shipped";
occurredAt: string;
};Outcome
- Three legacy services retired.
- Mean-time-to-resolve down from hours to minutes.
- Documentation got shorter, not longer — that's the tell.
If your stack feels like a Rube Goldberg machine, the answer usually isn't another tool. It's a sharper contract between the tools you already have.