Written and reviewed by Manish Kulkarni · About the author
Most people judge an AI tool by its first reply. If it is not quite right, they either give up or start over with a brand-new prompt. Both waste the work already done. The model has just shown you how it understood your request, and that is the most useful information you can get. Use it.
In the prompt lesson I asked ChatGPT for help planning a trip, and the reply was a list of six questions, not a plan. Answering them in one short line (dates, 8 days, 4 people, from Pune, ₹40,000, jungle safari) produced a full eight-day itinerary. That was already one round of iteration: the first answer told me what the model still needed, and my second message supplied it.
The plan ended with an offer to check safari availability for my dates and choose the best gate. That is the natural next round, and it shows the pattern: a good answer usually points to the next question.
Specific feedback gets you to a good result in two or three exchanges far more often than trying to write the perfect prompt on the first attempt. Three kinds of feedback work best:
Illustrative follow-ups for the trip plan (examples of what I would send, not output from a test): "Keep the dates but cut the budget to ₹35,000", "Replace the free day with one more safari", "Show the cost per person per day in a small table."
If the model has misunderstood the task itself, not just the details, patching it is slow. Open a fresh conversation, write one clear prompt that includes everything you have learned, and start from there. A good sign that you should restart is that your corrections keep fixing one thing while breaking another.
In October 2026 I took the 61-word About text that Claude had written for this site (see Prompting Fundamentals) and sent one short follow-up in the same chat: "Make it 30 words and mention that I work in data recovery."
The reply was exactly 30 words, and I counted them: "Learn AI Plainly explains AI in simple, honest language for beginners. I work in data recovery, and I'm sharing what I've learned so you can understand AI without the jargon."
What I noticed: one short message with two instructions was enough, because the first answer was already close. It did both things I asked: it cut the length and added my work in data recovery. Shortening also cost something. The friendly welcome and the line about AI not being an exclusive club were both dropped, and the text now opens with what the site does instead of a greeting. A follow-up that shrinks a text is worth reading against the original to see what disappeared. I would also check any line written in my own voice, such as "I'm sharing what I've learned", before publishing it.
For anything beyond a simple request, 2-4 rounds is typical for a genuinely polished result — don't expect perfection on the first try for nuanced tasks.
It can, but positive, specific direction ("do Y instead") tends to work more reliably than only describing what to avoid.
Sometimes — a fresh conversation with a clearer, more complete initial prompt can outperform continuing to patch a conversation that went off track early.