Part of the Level 1 core path · Lesson 2 of 5
The Refinement Loop
After this, you'll be able to read a Claude response critically and do one targeted follow-up that makes the output significantly better.
Before you start
Complete Ask Like a Real Human first; this lesson builds on writing a solid first prompt so your follow-up has something worth refining.
The idea
The first response is a draft, not a final answer. Claude gives you its best prediction based on what you wrote. Your job after that is to read it like an editor: what is right, what is off, what is missing. One focused follow-up beats five new prompts from scratch.
| Before | After | |
|---|---|---|
| Habit | Guess from a loose request | Use the lesson move |
| Work move | Skip The Refinement Loop | Apply The Refinement Loop |
| Check | No clear proof | Pass the lesson check |
The after column is the lesson target.
Most people skip the refinement step entirely. They accept a response that is 70% right and spend time fixing the other 30% themselves. Or they abandon the conversation and start over. Both are inefficient. The better move is to stay in the same conversation and give one specific correction.
Here is the before and after: Weak: 'Make it better.' Strong: 'This is good but the tone is too formal for our team. We are a casual startup. Rewrite the second paragraph to sound like a Slack message from a colleague, not a corporate memo.' The weak version gives Claude nothing to work with. The strong version gives Claude role (casual startup), format reference (Slack message), and scope (just the second paragraph). The revised output will be right.
The sequence that works every time: read the response and find the one thing that is most off. Name it specifically. Say what you want instead. That is one follow-up. Do not try to fix everything at once. Fix the biggest issue first, then decide if you need another round.
After two or three sessions doing this, the feedback loop becomes fast. You read, you name the issue, you get a better version. That loop is worth more than any single prompt technique.
Try it (12 min)
Watch out for
- Saying 'make it better' without saying what better means. Claude needs direction, not encouragement. Better for you might mean shorter, warmer, more specific, or more formal.
- Starting a brand new conversation every time something is off. Stay in the thread. The model already has your context loaded.
- Accepting something that is 70% right when one follow-up would get you to 95%. Most people leave value on the table by stopping one turn too early.
- Trying to fix everything in one follow-up. Pick the most important issue. Fix that. Then decide if you need another round.
Paste this into Claude
Here is what you gave me: [paste the part of Claude's response you want to improve]. This part is close, but [explain specifically what is off: too formal, too long, missing this detail, wrong tone, doesn't fit my audience]. Please revise just this section to [describe the adjustment you want].
What a good response looks like
Got it. Here's the revised second paragraph, rewritten to sound like a Slack message from a colleague rather than a corporate memo:
'Quick heads up on where we landed with the Thornton account. They pushed back on the timeline but were actually pretty reasonable about it once we walked them through the Q4 constraint. We've got until Nov 18 now instead of Nov 8. That gives the team a bit more breathing room. Priya is updating the project board today. If anything looks off once she's done, just ping her directly.'
The main changes: cut the passive voice ('it was determined that'), replaced 'the aforementioned timeline adjustment' with 'the timeline,' dropped the formal sign-off, and tightened the sentence about next steps so it sounds like something you'd actually say out loud. Want me to adjust the length or add anything about who needs to be looped in?What good looks like
- You read the first response and identified at least one thing that was close but not quite right
- Your follow-up message was specific about what to change, not just 'make it better'
- The revised response was noticeably closer to what you actually wanted
- You stayed in the same conversation thread instead of starting over
- You can describe in one sentence what made your follow-up effective
Go deeper (15 min)
Paste this into Claude
I want to practice the full refinement loop on a real task. Here is a task I need to do this week: [describe a real task, like 'write a bio for my company website' or 'draft talking points for a client call']. I am going to run three rounds with you. Round 1: give me your first attempt with no extra context from me. Round 2: I will give you one specific piece of feedback. Round 3: I will give you one more. After round 3, compare the three versions side by side and tell me: what changed with each round and what made each follow-up effective?
What a good response looks like
Here are the three versions of your company bio, side by side: Round 1 (no context): 'Sarah Chen is a seasoned marketing professional with over a decade of experience driving brand strategy and growth. She is passionate about connecting businesses with their audiences through innovative campaigns.' Generic, could be anyone. Round 2 (after your feedback: 'too corporate, I work with indie food brands'): 'Sarah Chen has spent 10 years helping small food brands find the words that make people stop scrolling. She works with indie makers, local producers, and scrappy food startups who know what they make is good, but struggle to say why anyone should care.' Better. Now it has specificity and voice. Round 3 (after your feedback: 'add something about my approach'): 'Sarah Chen helps indie food brands find their voice. She has spent 10 years working with small producers and local makers, turning product features into reasons people buy. Her approach: figure out what makes a customer choose you over everyone else, then say that, simply.' Usable. What changed each round: Round 2 added audience specificity. Round 3 added the 'how you work' signal that converts browsers into inquiries.
What good looks like
- You completed all three rounds without starting a new conversation
- Each follow-up was about one specific issue, not a general 'improve it' request
- The round 3 version is significantly closer to what you wanted than round 1
- You can name what changed between each round and why
When this breaks
- Breaks when the follow-up names a vague feeling ('I don't love it') instead of a specific change, because Claude cannot translate your taste into edits without a target it can aim at.
- Breaks when you start a fresh conversation after every imperfect response, because you throw away the accumulated context and force Claude to re-guess from scratch every time.
AI can help with this
After any response, tell Claude: 'Rate your own response on a scale of 1 to 10 for [clarity, tone, or usefulness, whichever matters most here]. What would you change if you rewrote it?' Use its answer as the basis for your next prompt. Let Claude do the self-critique so you do not have to.
You can now
Produce a follow-up message that names the specific issue, states the desired change, and yields a revised response noticeably closer to what you wanted.
Key takeaways
The refinement loop is the real skill. First prompt gets you started, second prompt gets you there. One specific follow-up is always worth it.
- Treat the first response as a draft. Read it like an editor and name the single thing that is most off.
- Strong follow-ups specify what to change and what to change it to, not just 'make it better.'
- Stay in the same thread. The accumulated context is what makes round two faster and sharper than round one.
- Fix the biggest issue first, then decide if another round is worth it. Do not try to repair everything at once.
Go deeper