After years of AI prompting, I’ve figured out the key question: when do you keep tweaking versus starting over? Here’s my framework for knowing which approach works.
The story
I’ve been elbow-deep in AI prompting for years now, and one question keeps coming up: when do you keep tweaking a prompt, and when do you just scrap it and start fresh?
Here’s what I’ve learned after three decades writing code and eight years building AI solutions.
Iterate when:
The output is almost there – maybe the tone’s off or it’s too wordy, but the core makes sense. The AI clearly gets what you’re asking for; it just needs better guardrails. A small tweak like “keep it under 200 words” fixes the problem.
Start over when:
The AI keeps misunderstanding you, no matter how you rephrase it. Your prompt has become a mess that you can barely parse yourself. The model latches onto the wrong pattern and won’t let go – usually because your initial framing was unclear.
The skills that matter most:
Be specific. AI doesn’t do well with vague. If you want something, spell it out.
Break big problems into smaller ones. Chain prompts. Keep each step manageable.
Watch your context window. When the model starts losing coherence, you’ve probably stuffed too much in.
Use examples. Few-shot prompting – showing the AI what you want – beats abstract instructions every time.
Get feedback. Your own eyes aren’t enough. Test with real users or teammates.
The real skill isn’t just knowing how to prompt – it’s developing an intuition for where the AI’s limits are, and when you’re fighting a losing battle. That intuition comes from repetition, from failing, from paying attention to what works.
Iteration and reset aren’t opposites. They’re two tools in the same kit. Know when to use each,
Recent Comments