Takeaways

(the wisdom extracted from months of pain)

• Know when to give up on a model — cut your losses early, but don’t overspend on a model you don’t need.
• Set MVP expectations realistically — revenue now is better than revenue later.
• Use deterministic code where possible — don’t reach for an LLM out of laziness.
• Use an LLM as a quality gate — combine deterministic checks with model‑based scoring.
• Version your prompts carefully — and remember: the surrounding code can change the answer too.

@ANTHROP\C — you can pay me later for free advertising.