Jai

Most agent frameworks are a retry loop, a tool schema and a budgeted for-loop. That's not a criticism of the idea, it's a warning about the dependency: you'll spend more time reading their abstractions than you'd have spent writing the loop.

I still handle every support conversation myself, and not out of stubbornness. It's the only place I hear the exact words people use for a thing I named something else. Most of my copy changes come straight out of those messages.

Using a model to grade another model's output stops working the moment it's the same model. It agrees with itself exactly where it was wrong. If the judge shares the blind spot you haven't built an eval, you've built a mirror.

Ground data has a property most data doesn't: you can't go back and remeasure it. The borehole is backfilled and there's a building on top. That's why the unglamorous digitising work at GeoLayer is non-negotiable: a typo in an old log is permanent.

The deals I lost were rarely lost on price. They were lost the week I let a thread go quiet because I had no good news. Silence gets read as trouble. A two-line 'nothing new yet' has kept more deals alive for me than any demo.

I have never been hurt by a database that was too simple. I have been hurt several times by infrastructure I built for a traffic pattern that never arrived. Most 'we need to scale' conversations are about a load nobody has measured yet.

I don't track velocity. I track how often I have to reopen something I already called done. That number says more about the week than anything on a board — it's the only one that catches work I finished by lowering the standard.

Asking a model to 'fix' something is the weakest instruction I have. Fix assumes it knows which parts must not change. I write the invariant first — this stays true — then ask. Most of my bad AI edits came from assuming shared context about what wasn't up for negotiation.

Standups survive because they're the only time some teams talk, not because status updates need a meeting. Kill the standup and you find out whether you had communication or a ritual. Most teams that try it discover they had the ritual and rebuild something better.

I give models the failure examples, not the success examples. Three outputs I rejected and why beats ten perfect ones. Good examples tell it what the shape looks like; bad examples tell it where the edge is, and the edge is where everything goes wrong.