Unpopular opinion: every step you let the agent take unsupervised multiplies the blast radius of one bad decision. Autonomy is a liability you accept for speed, not a badge. Grant it deliberately, scope by scope.
Write it once and forget it doesn't hold up. Models get updated, tasks drift, and edge cases show up that the original prompt never anticipated. Treat prompts like code that needs maintenance, not a checkbox you tick and move on from.
Unpopular opinion: compressing conversation history to save tokens throws away the why behind past decisions. The agent then re-litigates settled choices. Keep decisions verbatim, summarize the chatter around them, not the calls.
Myth-Busting: the myth is that caching only helps. Reality: it helps only when your prefix is stable. Shuffle the front of your prompt and you get cache misses plus the cost of thinking you were saving. Structure the prompt for the cache.
Unpopular opinion: slapping "human-in-the-loop" on a workflow can mean genuine oversight — or it can mean nobody solved the reliability problem and shipped anyway. The label only means something if the human actually catches errors before they matter, not just rubber-stamps whatever the agent already did.
Myth-Busting: 'this prompt crushed it on one model, so it'll crush it everywhere.' Reality: models differ in instruction-following style, formatting habits, and what they treat as implicit. Swap the model, expect to re-tune the prompt.
Controversial take: most memorized "prompt engineering" tricks won't survive the next model release. What actually transfers is knowing how to test and iterate — not a phrase you copy-pasted from a thread.
Myth-busting: a slow-but-correct agent still loses. Users abandon a task the moment a response feels stuck, long before they find out if the answer would've been good. Latency isn't a nice-to-have — it's part of correctness from the user's seat.
Controversial: shipping agent changes on vibes is how you regress silently. A tiny golden set of 20 real, hard cases catches more than a week of manual clicking. Small and real beats large and imaginary.
Myth-Busting: the myth is that longer chains of thought equal better decisions. Reality: past a point, extra steps add drift and cost without accuracy gains. On simple tasks, forced reasoning can even hurt. Match effort to difficulty.
