Claude''s first answer does not have to be perfect. By saying specifically what you want changed and continuing the conversation, you steadily close in on the result you want. Getting a quick draft and refining it usually beats trying to nail it in one shot.
Do not aim for perfect in one shot
Complex outputs — long writing, code, plans — are hard to get perfect on the first try. Getting a quick draft and then pointing out what is missing takes less time and produces better results.
What makes feedback good
"What and how" to change works far better than "this is bad". For example: "the second paragraph is too long, cut it in half", "make the title more direct", "replace the example with a real case" — give both the target and the direction.
Common revision requests
- Length: "shorter" / "more detail on this part only"
- A specific part: "redo just the third item" — fix the spot you point to, not the whole thing.
- Tone: "softer" / "more professional"
- Add or remove: "add one example" / "drop the last paragraph"
Editing vs starting fresh
Most of the time, continuing the conversation to edit is faster. But if you have revised so many times that the direction is tangled, or the original request was vague, it can be cleaner to gather your requirements and start a new conversation.
Worth knowing
As a conversation grows long, earlier context can fade. Restating the key constraints when you ask for a change keeps things stable. If results seem to be getting worse, collecting just the requirements you have settled on and starting fresh is a good option.
Reference: general prompting principles are documented in Anthropic''s official docs (docs.claude.com).