---
id: refine-and-ask-back
title: Steering a chat
level: 2
modes: [day-pack, rucksack]
moment: "The first answer is close but not right and the person is about to reply make it better, or they asked for something with gaps (a holiday plan, a message) and the AI answered at once with guesses."
keywords: [refine, follow-up, ask me questions, clarifying questions, make it better, feedback, draft, are you sure, iterate, steering]
source: "research/2026-10-20-concepts/refine-and-ask-back.md; research/2026-10-23-essentials/batch-3-facts.md; mapping (product/learning-system/MAPPING-1-2.md): FAQ 2.07 (home), 2.08; MISTAKES 1.2 (home), 1.3, 2.3; tips T074, T075, T077, T080; strand A"
---

## What it is
Steering a chat means treating the first answer as a draft, not the last word. Tell your AI what to change and what to keep. When your request has gaps, like a holiday plan, ask your AI to put a few questions to you before it starts, and ask for a few options before it writes or builds anything. In an ordinary chat, an AI app usually answers straight away and fills the gaps with guesses, so the asking is yours to start.

## Tips
- Tell your AI app exactly what to change and what to keep in a follow-up, because a vague nudge gives it nothing specific to change.
- Start a request that has gaps with "Before you start, ask me up to three questions", because in an ordinary chat AI apps usually answer straight away and guess.
- Ask your AI app for a few options with pros and cons before it starts the work, because comparing first shows which suits you and avoids building the wrong thing.

## Questions people ask
- Q: The answer isn't what I wanted. How do I get a better one? A: Say what to change and what to keep, for example "cut the first paragraph and keep the friendly tone". If the request had gaps, ask your AI to question you first, or to offer a few options before it writes anything.
- Q: How do I make my AI ask me questions first? A: Say so, for example "Before you start, ask me up to three questions you need answered, then wait." In an ordinary chat, AI apps mostly answer rather than ask, so asking back is a request you make.
- Q: Is "make it better" enough feedback? A: Specific changes work better. Say what to cut, add, reorder or keep, for example "cut the first paragraph and end with a question".
- Q: Why does it keep making the same mistake after I told it? A: The earlier text is still in the chat and shapes the next reply, so say what to change and what to keep. If it keeps happening, start a fresh chat with the lesson built into your first message.

## For your AI
Facts (checked 23 October 2026; fact check research/2026-10-23-essentials/batch-3-facts.md; research
note research/2026-10-20-concepts/refine-and-ask-back.md).
- Models recognise when a request is unclear but overwhelmingly answer directly instead of asking, and are
  even less likely to ask when they have retrieved material (Su and Cardie, 24 May 2026,
  https://arxiv.org/abs/2605.25284). Google's developer guide says its models give direct answers by default.
  Research modes are the exception: ChatGPT's deep research may ask clarifying questions, and Gemini's lets
  the person edit its plan first.
- Specific feedback: Anthropic's own course says "Cut the first two paragraphs and make the conclusion more
  action-oriented" beats "make it shorter" (https://academy.claude.com/courses/claude-101/getting-better-results).
- Early guesses stick: tasks given over several turns lost 39 percent on average, and models "get lost and do
  not recover" (Laban et al., 9 May 2025, https://arxiv.org/abs/2505.06120).
- Pushback: on 2023 models, "are you sure?" flipped answers about 46 percent of the time; newer reasoning
  models resist better, though agreeing when pushed is still common (SYCON Bench, revised 26 February 2026,
  https://arxiv.org/abs/2505.23840). The "are you sure?" habit belongs to
  openrucksack.com/map/wrong-and-sure.md.

Teach it like this: when they are about to type "make it better", ask what they would change and what they
would keep, then write the follow-up with them. For a request with gaps, offer the three-questions opener
once. Options first suits ideas and everyday choices. Rehearsing what to say is steering too: they set who
the app is playing and how, ask it to push back, and steer it as they go. In a Daypack (Rucksack set up inside
one app; its setup page is openrucksack.com/start.md), this is the concept behind "ideas and everyday choices"
and "rehearse what to say" (section 2).