Talk, don't search: following up and improving an answer
Get from "not quite right" to "exactly right" by having a conversation instead of starting again.
Search engines taught us to type a few words and accept what comes back. AI works best the opposite way: like a conversation.
The first answer is a starting point. Then you shape it:
- "Shorter."
- "Simpler, please. I don't know what 'equity' means."
- "Good, but make it sound less like a robot."
- "Give me an example."
- "Can you put that in a table?"
- "Why did you suggest that?"
The assistant remembers everything said earlier in the same chat, so you don't need to repeat yourself.
Two tips:
- Say what's wrong, not just "try again". "Too formal" is far more helpful than "no".
- Start a new chat for a new topic. Old topics can confuse new answers.
People who get the most out of AI usually go back and forth three or four times. That's not failure; that's how it's meant to work.
The biggest habit change
With search, you type, scan the results and leave. With an AI assistant, the real value comes from the follow-up. Each message builds on everything before it in that chat, so you can steer the answer step by step.
Useful follow-ups to keep in your back pocket
| To... | Say... |
|---|---|
| Change length | "Half the length." "Expand the second point." |
| Change tone | "Warmer." "More professional." "Less salesy." |
| Change level | "Explain it like I'm new to this." "Assume I know the basics." |
| Change shape | "As a table." "As a checklist." "As three bullet points." |
| Get more | "Give me five more options." "What did you leave out?" |
| Understand | "Why do you suggest that?" "What does that word mean?" |
| Challenge | "What are the downsides?" "What would someone who disagrees say?" |
| Fix | "That date's wrong, it's the 14th. Update the whole thing." |
Be specific about what's wrong
"Try again" leaves the assistant guessing. "The second paragraph sounds like an advert; rewrite it so it sounds like a normal person" tells it exactly what to change. Quote the bit you don't like if that's easier.
Keep what works
When you ask for a change, it may rewrite things you liked. Say "keep everything else the same" or "only change the opening line".
When to start a new chat
- New topic. A fresh chat avoids old details leaking into new answers.
- The conversation has gone in circles. Start again with a better first request, using what you've learned.
- It's very long. Assistants can lose track of early details in very long chats. Start a new one with a short summary: "Here's where we got to: ..."
Use the edit and retry buttons
Most apps let you edit your last message or regenerate an answer. Editing is often faster than adding a correction, because the assistant then answers the improved question cleanly.
Why conversation works so well
Each time you send a message, the assistant reads the whole conversation so far before replying. Your follow-ups therefore act as extra context and correction, steadily narrowing in on what you want. That's very different from a search engine, which treats every query as new.
Conversation length and the context window
There's a limit to how much an assistant can take into account at once, called the . Modern assistants have large windows (often hundreds of pages of text), but quality can still slip in very long chats: early details may get less attention, or the assistant may repeat itself. Signs of this include forgetting a constraint you set at the start or contradicting an earlier answer.
Practical fixes:
- Restate key constraints when they matter: "Remember: vegetarian, under 30 minutes."
- Summarise and restart: ask "Summarise the key decisions so far in five bullets", then paste that into a fresh chat.
- Use projects or custom instructions (lesson 14) for things that should apply to every chat.
Steering techniques
Iterative refinement. Draft, critique, revise. You can even ask the assistant to do the middle step: "List three ways this could be better, then rewrite it with those improvements."
Branching. Editing an earlier message creates a new branch of the conversation. It's a clean way to try a different direction without the assistant being influenced by the path you abandoned.
Anchoring. If you like one part, anchor to it: "The tone of paragraph 2 is perfect. Make the rest match it."
Perspective shifts. "Now read this as my manager would. What would worry her?" or "Explain the opposite view as convincingly as you can." These help you stress-test ideas, not just polish words.
Fact correction. When you correct a fact, ask the assistant to update everything that depended on it. Otherwise the wrong fact can linger elsewhere in the text.
Knowing when to stop
Endless tweaking has diminishing returns. If after four or five rounds it's still not right, the first request was probably missing something important. Start again with a better first request; you'll usually get there faster.
Getting a second opinion
Different assistants have different strengths. For important writing or decisions, paste the result into a different assistant and ask for a critique. It's a quick way to catch blind spots.
Try this prompt
That's a good start. Now [make it shorter / make it friendlier / add an example / explain the second point more simply]. Keep everything else the same.
Replace the [brackets] with your own details
On haystak
Quick check · 1 of 3
The answer is too formal. What's the best follow-up?
Next step
Sharing files: documents, photos and spreadsheets
Use AI to read, explain and work with your files, from a 40-page PDF to a photo of a letter.