Five habits that work on every model
Before you tune for a specific model, get these right. They hold across Claude, ChatGPT, Gemini, and Kimi — and most "the AI gave me junk" moments trace back to skipping one of them.
Be explicit — say exactly what you want
Models are not mind-readers. State the audience, the format, the length, and what 'done' looks like. 'Summarise this for a skeptical CFO in 5 bullets, max 15 words each' beats 'summarise this' every time — and it's cheaper, because you don't burn a second round fixing the shape.
Show an example of the output you want
One good example of the format or tone you expect is worth a paragraph of description. Paste a sample row, a sample sentence, or a "good vs bad" pair. The model will match it far more reliably than it will follow an abstract instruction.
Structure your context — separate it from your ask
Keep the material (the transcript, the PRD, the data) visually separate from the instruction. Headings, dashes, or XML-style tags all work. When the model can tell "here is the input" from "here is the task," it stops blending them — and stops re-reading, which saves tokens.
Match the model — and the tier — to the job
Don't reach for the flagship by reflex. Rote extraction and tagging are a fast-tier job (Haiku, Flash, Luna). Genuine reasoning and trade-offs are a flagship or reasoning-tier job. Picking the right tier is the single biggest lever on cost for repetitive work.
Iterate — your first prompt is a draft
Treat prompting as a short conversation, not a one-shot. If the output is 80% right, say what's off ('too long', 'too generic', 'lead with the risk') rather than rewriting from scratch. Steering an existing answer is faster and cheaper than starting over.