Refine, don't restart — the second draft is where it lands
The biggest skill jump at this level: stop deleting the conversation and starting over when an answer is wrong. Refine in-place.
::TL;DR · the whole lesson in three lines
- MOVEThe biggest skill jump at this level: stop deleting the conversation and starting over when an answer is wrong. Refine in-place.
- DRILLTake any prompt that produced an OK-but-not-great answer this week. Refine it in three turns without starting over.
- WINYou sent four messages in one chat without restarting.
::concept · what's actually happening
Novice users hit a wrong answer and start a new chat. Learner users push back on the wrong answer in the same chat. The second draft is almost always better than the first.
read full concept · 2 more paragraphs →collapse concept ↑
Useful follow-ups: "That's too long, cut it in half." "You missed [X] — rewrite incorporating it." "The tone is too formal — make it sound like a friend wrote it." "You assumed [thing] — what if [other thing]?" "Show me 3 versions, varying [dimension]."
The conversation has memory inside the conversation. The AI remembers what it just said and what you just corrected. Use that.
::drill · do the thing
Take any prompt that produced an OK-but-not-great answer this week. Refine it in three turns without starting over.
::L4 drill · copy-paste into any AI chat
(Re-use the original prompt that disappointed you. Then push back across three turns.) Turn 2: "This answer is [specific complaint — too long / too formal / missed X / wrong tone]. Rewrite, fixing that. Keep [what was good]." Turn 3: "Better. Now [next thing to fix]. Same constraint." Turn 4: "One more pass. [Final polish — make it tighter / make it sound like me / cut the last paragraph]."
::steps
- 01Open the chat where you got a meh answer recently.
- 02Don't restart. Use Turn 2 above.
- 03Use Turn 3.
- 04Use Turn 4.
- 05Notice how the answer landed by turn 4 vs. where you started.
::worked example · what one real run looks like
See one real run before you try yours.
::worked example · what one real run looks like
See one real run before you try yours.
::what a real human filled in
I asked Claude: "Write a 200-word LinkedIn post announcing my new freelance copywriting service for SaaS startups." It gave me a decent post — clear, professional, hit all the basics (services offered, target audience, CTA at end). But it reads like every other freelancer launch post on LinkedIn. Generic. Could've been written by anyone. I want to push it across 3 turns in the same chat to get something that actually sounds like me and stops people scrolling.
::what the AI returned
Good drill — this is exactly the kind of "OK but forgettable" output that benefits most from refinement instead of restart. Here's how the 3-turn push would look in practice.
**Turn 1 (you push back):** "This is structurally fine but it reads like every freelancer launch post on LinkedIn. Don't restart — just diagnose: what specifically is generic about it? Name 3 phrases that any of 10,000 other freelancers could have written."
What good looks like in the response: Claude should pull actual phrases from the draft — things like "passionate about helping startups grow" or "let's connect" or "specializing in conversion-focused copy" — and explain why each is filler. If it just gives you abstract critique ("it's too generic"), push harder.
**Turn 2 (you redirect the energy):** "Right. Rewrite, but this time: open with a specific failure I've seen in SaaS copy (you can invent a plausible one — bad onboarding email, dead landing page header, whatever). No 'excited to announce.' No 'passionate.' No CTA that says 'DM me.' The post should make a SaaS founder think 'this person has actually read my homepage and it's not great.'"
This is the load-bearing turn. You're not just saying "make it better" — you're banning specific failure modes and replacing them with a concrete creative constraint (lead with an observation, not an announcement).
**Turn 3 (you tighten):** "Closer. The opening works. But the middle still drifts into generic services-offered territory in the 4th sentence. Cut that sentence entirely and let the specificity of the opening carry the credibility. End with a line that sounds like something I'd say to a friend at a bar, not a CTA."
By turn 3 you should have something that sounds like a person, not a template. If it still doesn't, the issue is usually that you haven't given Claude enough of your actual voice to mimic — at that point, paste in 2-3 sentences of your real writing (a Slack message, an email, a tweet) and ask it to match the cadence.
**One thing to watch for:** you'll be tempted to restart on turn 2 because the first refinement won't feel like a big enough jump. Don't. The compounding happens on turn 3.::what to notice
- Turn 1 doesn't ask for a rewrite yet — it asks Claude to 'diagnose: what specifically is generic about it' and 'name 3 phrases.' Diagnosis before rewrite is the move that makes refinement work; restarting skips this and loses the signal.
- Turn 2 bans specific failure modes ('No excited to announce. No passionate. No CTA that says DM me.') instead of giving vague direction. Notice how the constraint is negative + concrete, not 'make it punchier.'
- The example flags turn 2 as 'the load-bearing turn' — a small piece of meta-commentary that tells the user where the work actually happens, so they don't bail too early.
- The 'one thing to watch for' note at the end addresses the exact moment a human would abandon the drill: 'you'll be tempted to restart on turn 2 because the first refinement won't feel like a big enough jump.' That's the failure mode this lesson exists to prevent.
- Small imperfection: the line 'What good looks like in the response' is structurally awkward — it's a hedge about what Claude should do, sitting inside an example of what the human should say. A real AI response would have this slight tonal wobble; a polished version would smooth it out.
::outcome · what should be true
- You sent four messages in one chat without restarting.
- Turn 4 is meaningfully better than Turn 1.
- You internalize: the prompt is not the lesson. The conversation is.
::trap · the most common failure
Refining forever. After 5–6 turns, if it's not landing, the original prompt was wrong. Start fresh with a better prompt, using what you learned.
::other lessons at Learner level
The verify rule — three categories of trust
Not everything AI says needs verification. Most things don't. Knowing which third does is the skill.
Your saved-prompt library — the second-biggest leverage
The first time you write a good prompt for a recurring task, save it. The second time, you reuse it. By month two, your prompt library is doing 60% of the work.
Refusal posture — knowing what your AI won't say
Every AI refuses different things in different ways. Map the refusal shape of the tool you actually use, instead of guessing or repeating internet rumors.
Few-shot — teach by example
Three good examples will outperform a one-paragraph instruction every time. The skill is curating the examples.
::part of the AtomEons /learn curriculum · 45 lessons · 5 levels · cc-by 4.0