πŸŽ“ Courses
AI Literacy Course
🧠 Module 1: Reverse Prompting
Iteration: Refining vs. Starting Over

Iteration: Refining vs. Starting Over

Iteration: Refining vs. Starting Over
Module 1: Reverse Prompting Β· Lesson 5

Module 1, Reverse Prompting | Essay 5 of 8


When an output isn't right, most people delete it and start again. New prompt, blank slate, try a different angle. It feels productive. It's usually the wrong move.

Starting over throws away something valuable: the context that's already built up in the conversation. The agent knows what you've been working on. It understands the direction you're heading. That context is working in your favor, even when the last output missed. The moment you start a new conversation, it's gone.

The better move almost always is to stay in the conversation and tell the agent what's wrong.

Not "make it better." That's not feedback, it's a wish. Useful feedback is specific and directional. "The tone is too formal, I need this to sound like me, not a press release." "The second paragraph repeats what you said in the first. Cut it." "You buried the main point. Lead with it."

Three questions that sharpen any feedback:

  • What's wrong with it? (specific, not general)
  • What would right look like? (even a rough description helps)
  • What do you want to keep? (so the agent doesn't fix the broken part by breaking the parts that worked)

That third one matters more than people think. If you tell an agent "this isn't right" without saying what to preserve, it'll rewrite everything. Sometimes that's what you want. Usually it isn't.

So when do you actually start over? Two situations. First: the output went so far in the wrong direction that continuing would compound the confusion. If you asked for a two-paragraph summary and got a 10-section report, the conversation has the wrong frame. Start fresh with clearer constraints. Second: the tone is fundamentally off in a way that's hard to correct incrementally. Voice mismatches are easier to fix at the start than to edit out later.

Everything else: iterate. Stay in the conversation. Be specific about what's wrong and what you want instead.

The agents you'll get the most from aren't the ones you prompt perfectly on the first try. They're the ones you learn to have a real back-and-forth with.

πŸ‹οΈ

Pod Exercise: Take the email output from Essay 1's exercise, the generic one from the vague prompt. Instead of rewriting the prompt, reply to the agent in the same conversation with specific feedback about what's wrong. Run two or three rounds of refinement. See how close you can get to something you'd actually send.