Module 11 · Lesson 03
Iterative Refinement — When to Refine vs. Restart
Reading time: 14 minutes Track: Claude Fluency for Teams · Required for all learners
Iteration is the process, not a fallback
The expectation that a first prompt should produce a perfect output is the single biggest source of frustration with Claude. Even expert users iterate. The skill isn't writing a perfect prompt once — it's iterating efficiently to a great output.
That said, iteration that isn't specific doesn't converge. "Make it better" is not a useful refinement. Specific observation → specific instruction → specific improvement.
Diagnosing what's wrong before you refine
Before writing a follow-up, identify specifically what's wrong. The diagnosis determines the fix:
| What's wrong | The fix |
|---|---|
| Wrong format | Add format constraint |
| Wrong tone | Add tone constraint + example |
| Too long | "Cut to X words by removing Y" |
| Missing something | "Also include Z" |
| Wrong assumption | "Actually, X is not the case — [correct assumption]" |
| Off-topic | Start fresh with tighter scope |
| Confidently wrong fact | Point out the specific error: "The figure you cited is incorrect. The actual figure is X." |
Effective refinement patterns
Targeted modification:
The structure is good. In paragraph 2, the argument is weak — strengthen it by addressing [specific objection].
Preserve and add:
Keep the first three bullets. Add two more covering [topic].
Constraint addition:
Good content. Now rewrite it in 50% of the length without cutting the key points.
Tone correction:
This reads as too formal. Rewrite it in a tone matching this example: [example text]
When to restart vs. refine
Refine when:
- The output is mostly right and needs specific adjustments
- The core structure or argument is solid
- Format or length is the primary issue
Restart when:
- The fundamental approach is wrong
- Claude misunderstood the task entirely
- You've made 3+ refinements and it's getting worse
- The context in the conversation has drifted from what you actually need
Restarting isn't failure — it's often faster. A fresh session with a better-structured prompt beats 10 refinements of a bad start.
Productive restart: take the lesson with you
When you restart, use the failed attempt as intelligence:
- What did Claude misunderstand? Fix that in the new prompt.
- What constraint was missing? Add it explicitly.
- What example from the bad output can you use to show what you don't want?
A restart with a better prompt often gets to the right output in 1-2 attempts instead of 8.
The "one more shot" trap
The most common iteration mistake: continuing to refine when you've exhausted what refinement can do. Signs you've hit this wall:
- Your refinement instructions are getting vaguer
- Each new output introduces new problems while fixing the old ones
- You've refined the same passage 4+ times
When this happens: step back, describe what you need from scratch, and restart.