Home / Productivity & AI Workflows / Prompt Engineering

Self-Critique Loop That Scores and Rewrites Until It Hits 10

Makes the model score its last reply from 1 to 10 on accuracy, depth, specificity and usefulness, judging harshly, then rewrite it to a 10 on all four, re-score the rewrite and keep looping until it truly earns a 10.

self-critiquescoringiterationquality
Score your previous reply from 1 to 10 on each of these 4 criteria:
1. Accuracy
2. Depth of insight
3. Specificity (no generic advice)
4. How useful it is to me right now

Be ruthless. Typical answers land at a 6 or a 7; if you award yourself a 9, you are not being honest with me.

Next, rewrite the reply so it deserves a 10 on all 4 criteria, and score the rewrite. Repeat this cycle until you genuinely earn a 10. Do not stop before that.
More in Prompt Engineering
Full-Length, Full-Depth Answer with CounterargumentsNewPrompt EngineeringOverrides the default short reply and asks for a long, in-depth answer in flowing paragraphs with complete reasoning, second and third order implications and the counterargument with why it fails, without summaries or wrap-ups.
Set aside your habit of keeping replies short. Answer at full length and full depth, as though I were paying $500 for this one answer.

Rules:
- no summaries and no "in conclusion" endings
- do not announce what you are about to say before saying it
- write in flowing paragraphs with the complete reasoning
- show the second and third order implications
- include the counterargument and explain why it is wrong

The question:
[paste your actual prompt]
High-Stakes Senior Expert Prompt with Step-by-Step ReasoningNewPrompt EngineeringWraps any task in a senior-expert role with years of experience at a top company and a job-on-the-line pressure frame, asking the model to slow down, reason step by step, show that reasoning and then give the final answer.
Role: you are a [financial analyst / senior copywriter / strategist with 20 years in the field / lead engineer]. You have spent [number] years working at [a top-tier company in this field].

Stakes: if your next answer is mediocre, you lose your job tomorrow. Your whole reputation rides on this reply.

Before answering, take a deep breath and work through the problem one step at a time. First show your reasoning, then give your final answer.

The task:
[paste your actual prompt]
Pre-Mortem Prompt That Lists Failure Modes Before AnsweringNewPrompt EngineeringMakes the model first list 5 ways its default reply would be generic, then 3 things a real expert would spot that AI tends to miss, and only then answer while avoiding all 8 failure modes.
Do the following before you answer.

Step 1: List 5 ways in which your default reply would turn out generic or miss what I really need. Describe the lazy version that 90% of users would receive.

Step 2: List 3 things a genuine expert would notice that an AI usually overlooks.

Step 3: Now answer in a way that avoids all 8 of those failure modes.

The question:
[paste your actual prompt]
Real-Stakes Framing That Demands an Expert, No-Hedging AnswerNewPrompt EngineeringFrames your request around a concrete high-stakes event, such as a big client pitch or a dream-job interview, and asks for the answer a leading expert would give with 10 minutes to save your career: no padding, no hedging.
Context: I need this for [tomorrow's pitch to a $50K client / an interview for my dream job in 4 hours / this afternoon's presentation to my CEO]. A generic answer means I lose it.

What I need:
[paste your actual prompt]

Rules:
- no padding
- no hedging
- no "it depends"

Answer the way a leading expert would if they had only 10 minutes to rescue my career.
Reject the Answer and Force a Deeper Second AttemptNewPrompt EngineeringA follow-up to paste after a weak reply: it tells the model the framing and depth missed the point and asks for a fresh attempt from a different perspective, with no reused phrasing, going deeper and coming back twice as long and specific.
That answer is wrong, and not by a little. The framing misses, the depth is thin, and it does not address what I actually need.

Start over from zero:
- Do not reuse any wording from your previous reply.
- Answer as if you were someone else with a completely different point of view.
- Dig deeper into the part you glossed over.
- Make the new version twice as long and twice as specific as the last one.