1.2Iterate & refine an answer
The essentials in 30 seconds
- Never abandon a 70% answer: three short iterations (“shorter”, “different tone”, “add X”) bring it close to final.
- Ask for variants: “suggest 3 versions” costs a few seconds and gives you choice instead of imposed output.
- Bogged-down thread = fresh thread: after too many contradictory back-and-forths, restart a clean conversation with a consolidated prompt.
Threaded mini-case — Amine sculpts his invitation in Algiers
Amine requests an invitation to the quarterly progress meeting and receives a long, flat text. Instead of retyping everything, he iterates in the same thread: “Shorter, 80 words max”, then “More enthusiastic”, then “Add the agenda in 3 bullets”. He then asks for two variants (sober and dynamic), chooses, then rebuilds into ONE complete RCTF prompt the request that would have produced this result directly.
Sequel to case 1.1 — Amine refines instead of restarting.
1. Iterate, don't retype
Beginners throw away the first imperfect answer and retype everything; effective users iterate: “shorter”, “more formal”, “add a quantified example”, “keep the structure but change the tone”.
The conversation keeps the context — each instruction refines the previous result, as with a colleague. Amine keeps his invitation in the same thread and sculpts it in small steps.
2. Variants and self-critique
Refinement instructions serve every day (length, tone, structure, angle); requests for variants (“suggest 3 versions: neutral, warm, direct”) exploit the model's variability. And guided self-critique (“reread your answer like a demanding director, then apply your corrections”) markedly improves deliverables.
Amine requests two tones for his invitation and has the AI itself proofread the text before choosing: he compares instead of accepting.
3. Escape route and consolidation
When the thread gets bogged down after too many contradictory back-and-forths, restart a clean conversation with a rebuilt prompt. Then consolidate: rebuild into ONE RCTF prompt the request that would have produced the final version directly — that is how you progress.
Amine notes his consolidated prompt in his 1.1 template: the next invitation will start directly at the right level.
The iteration blooper reel (never do this)
Moral: the thread keeps the memory; you keep the course.
🔎 Concrete example — 🔎 Concrete example — From 70% to 95% in 3 iterations
Don't abandon an imperfect answer. Test the sequence as it is.
“Write an invitation message for the quarterly team meeting”
→ “Shorter, 80 words max” → “More enthusiastic” → “Add the agenda in 3 bullets”
→ “Suggest 2 versions: one sober, one more dynamic”
→ consolidate into ONE complete RCTF promptReading: each short instruction refines the previous result without starting from scratch.
Lesson: asking for variants costs a few seconds and gives you choice.
Section vocabulary
- Iteration
- Refining the answer through successive instructions instead of reformulating the whole request.
Simple example: Amine's three touch-ups on his invitation (length, tone, bullets).
What it is not: throwing away the answer and retyping the whole request each time.
Reflex: small successive steps in the same thread.
- Variants
- “Give 3 different versions”: exploiting the model's variability to choose rather than accept.
Simple example: The sober and dynamic versions Amine compares before sending.
What it is not: accepting the first suggestion with no alternative.
Reflex: always ask for at least two versions for a visible document.
- Bogged-down thread
- After too many contradictory back-and-forths, answers degrade: you must restart a fresh thread.
Simple example: The confused thread Amine abandons for a consolidated prompt.
What it is not: persisting ten more messages on the same thread.
Reflex: beyond 4-5 back-and-forths: fresh thread.
- Guided self-critique
- Asking the AI to reread and correct its own answer with a demanding eye.
Simple example: “Reread this invitation like a demanding director, then apply your corrections”.
What it is not: proofreading alone without ever having the text challenged.
Reflex: have it critiqued before validating a deliverable.
In practice — In practice — Sculpt an answer in 4 steps
- Ask: “Write an invitation message for the quarterly team meeting” (deliberately loosely framed).
- Iterate: “Shorter, 80 words max” then “More enthusiastic” then “Add the agenda in 3 bullets”.
- Ask for variants: “Suggest 2 versions: one sober, one more fun” — choose.
- Consolidation exercise: rebuild into ONE RCTF prompt the request that would have produced the final version directly — that is how you progress.
Ready-to-say script — coach a colleague who retypes everything (30 s)
“Don't retype everything: keep the thread and give short orders — shorter, more formal, add the agenda in bullets. Ask for two variants and choose. If the thread gets tangled after 4-5 attempts, restart a fresh conversation with everything you learned in a single prompt.”
Say it in front of the colleague's screen, on their real case.
Copy-paste model — iteration grid
STARTING PROMPT: _______________
ITERATION 1 (length): _______________
ITERATION 2 (tone): _______________
ITERATION 3 (content/structure): _______________
VARIANTS: sober [ ] | dynamic [ ] | choice: _______________
CONSOLIDATED PROMPT (RCTF): _______________Almost-good answer
This is the ideal case: two or three short instructions are enough — retype nothing, refine in the thread.
Amine's team in Algiers
Amine shares his iteration grid with his team: everyone consolidates their RCTF prompts after use — the quality of project documents rises within weeks.
Thread going in circles
If each correction breaks another: stop, fresh thread, consolidated prompt with all the context and explicit format.
Sensitive deliverable
For a binding document: iterate on substance AND have the AI proofread it (“what is missing?”) — but final validation stays human.
✓ Best practices
- Iterate, don't retype.
- Variants to choose from.
- Consolidate into RCTF.
✗ Common pitfalls
- Retyping everything.
- Persisting on a bogged-down thread.
- Never capitalising.
Key takeaways
- Never abandon a 70% answer: three iterations bring it close to final.
- Short instructions are enough: the conversation keeps the context.
- Asking for several variants costs a few seconds and gives you choice.
- Confused thread = fresh thread with consolidated prompt: don't persist.
48h challenge — apply tomorrow
Take one real document (invitation, announcement) and sculpt it in 3 iterations + 2 variants (grid above), then note your consolidated prompt. 15 minutes is enough.
- I can name the 3 moves (refine, variants, fresh thread).
- I know when to restart a clean conversation.
- I sculpted a real document + noted my consolidated prompt.
Frequently asked questions
How many iterations before it becomes wasted time?
Beyond 4-5 back-and-forths on the same point, the initial prompt was poorly framed: rebuild it. Iteration refines; it does not repair absent framing.
Can I ask the AI to critique its own answer?
Yes, and it is very effective: “Reread your answer like a demanding director: what is missing?” — then “apply your own corrections”. Guided self-critique markedly improves deliverables.
What do effective users do with an imperfect answer?