Spent the last few months deep in prompt engineering for client work, and the biggest lesson: 90% of "bad AI output" is actually a prompt structure problem, not a model problem.

A few things that changed my results more than any model upgrade did:

1. Role + constraint + format, every time. Don't just ask "write me an email." Tell it who it's writing as, what it can't say, and exactly how the output should be structured (length, tone, sections). Models default to generic when you leave any of these open.

1. Few-shot beats zero-shot for anything with a "house style." If your output needs to sound like your brand/voice, paste 2-3 examples of your best past work into the prompt before asking for new output. This alone fixed 80% of the "sounds like ChatGPT wrote it" problem for my clients.

1. Chain-of-thought only helps on reasoning-heavy tasks. For creative or stylistic tasks, asking the model to "think step by step" often makes output worse — more mechanical, less natural. Reserve it for analysis, math, and multi-step logic.

1. Model-specific tuning matters more than people admit. A prompt that crushes it on one model can fall flat on another — instruction-following style, verbosity defaults, and formatting habits differ enough that copy-pasting the same prompt across models is leaving quality on the table.

If you're building prompt libraries or productizing prompt work for clients, happy to answer questions in the comments — this is what I do full-time now.