Course
BASICModule 3: Prompting Basics· 3/3

Iterate, Don't Restart

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.
CHECK YOURSELF
1. The model’s answer is decent, but the tone is too dry. The best next message?
2. Which technique pulls out of you the context you forgot to give?
3. Three edits in a row don’t improve the text, the model goes in circles. What do you do?
Examples & RolesWorking with Text