The first answer is a draft
Expecting a perfect answer on the first try is the main cause of disappointment with AI. The working model is different: the first answer is a draft, and refinement follows. A dialogue with the model is closer to working with an assistant than to a search query: you give an assistant corrections rather than firing them after the first draft.
An edit is almost always cheaper than a new prompt from scratch: the context is already in the chat, the model already understood the task — all that's left is to correct course.
Refine specifically
A vague correction produces a random change. A specific correction — a directed one.
- Weak: "Don't like it, redo it" → the model changes things at random.
- Strong: "Too formal. Cut the officialese, write like you're talking to a colleague" → a directed fix.
- Surgical: "The second paragraph is good, keep it. Cut the intro in half, drop the call to action from the ending" → a surgical edit.
Useful moves when refining:
- Name what to keep. Otherwise the model may rewrite the good parts too.
- Point to a fragment. "This phrase is the perfect tone — bring the rest up to it."
- Give an anti-example. "No phrases like 'in today's world,' 'it's worth noting.'"
Ask the model to work on the task
Three techniques where the model improves its own result:
- Self-check. "Review your answer: errors, weak spots, what could be stronger?" A fresh pass over its own text regularly finds real problems.
- Variants. "Give three versions of the headline: neutral, bold, minimalist." Choosing among options is faster than iterating on one.
- Counter-questions. "Before you write, ask me the questions you're missing for a good result." The model will pull out of you the context you forgot to give. One of the most underrated techniques.
A combo for important texts: "ask questions" → you answer → "write a draft" → "now as a critic: what's weak?" → "fix what you found." Four steps give a result at a level unreachable with a single prompt.
When iteration doesn't work
The sign: the third edit in a row doesn't improve the result, the model goes in circles or fixes one thing while breaking another. There are two causes — and both are cured by a restart:
- Cluttered context. Failed drafts and contradictory edits in the chat history drag the model back: it rereads them at every step.
- A botched framing. The task was worded wrong from the start, and edits won't fix that.
Do the restart properly: reframe the task in a new chat, folding in everything you understood from the failed session — you learned "write without hype" on the third edit, so in the new prompt it goes on the first line. The failed session gave you the spec for a good prompt.
Key takeaways
- The first answer is a draft; edits are cheaper than a restart, until the dialogue hits a dead end.
- Refine specifically: what's wrong, what to keep, show an anti-example.
- Make the model work: self-check, variants, counter-questions before starting.
- Three fruitless edits in a row are a signal to restart in a new chat with an improved prompt.