On this page
Quick notes on what I picked up from a client today.
Want AI working in your own job, not just in demos? See the 3-month 1-on-1 program →
The client’s root causes
- No spec settled, conclusions never written to a file, context too messy, long-context model output drifts.
- Wrong model tier and effort setting, so the model has no budget left for verification.
- No rhythm of diverging and converging: the model converges when you want to extend, and diverges when you want to close in.
What I prescribed
1. Three-tier model split: Fable plans, Opus executes, Sonnet investigates, paired with effort settings.
Since we’re on model pairing, here’s one I do backwards. Most people seem to pair their models as “strong model for the main agent, weak model for the subagents.” Lately I’ve also come to like the other way round: “strong model for the subagent, weak model for the main agent.”
I think of it as hiring a consultant for a one-off piece of advice. It works well, and when the scope is clear (you name the docs it should read) it saves a lot of tokens.
A small temple doesn’t have to keep a big Buddha on the altar year-round.
2. Use grill-me to press vague feelings into a concrete plan: known-knowns / known-unknowns / unknown-unknowns.
3. How to write a diagnostic prompt: ask the AI to pinpoint the root cause first, then build traceable/observable infrastructure (a log at every node).
4. Four tools: explain (re-explain in plain language), history-find (semantic search across tools), stt (voice input), Codex plugin (adversarial review).
Model governance: only a hook can cure rambling
While helping the client sort out prompts and rules, I confirmed one thing all over again: only a good hook can cure Opus 5’s rambling, just so you know. The official advice is that rules for the 5-series models should be lean, but the 5-series models themselves have no idea about this.