alivirgo Give AI Direct Prescriptive Feedback (AI Skill)
Overview
When an AI response doesn't hit the mark, many users respond with vague frustration ("No, that's not good, try again" or "Make it more creative"). Because the model cannot read your mind, it randomly varies tokens, often worsening the output or abandoning the parts that were already working.
The Direct Prescriptive Feedback Protocol teaches users how to steer subsequent conversation turns using the 3-Part Feedback Formula: validate what worked, isolate the flaw, and provide an explicit structural fix.
Vague Complaint vs. Prescriptive Steering
┌─────────────────────────────────────────────────────────────┐
│ Feedback Loop Comparison │
│ │
│ Vague Complaint: │
│ "Make this sound better and less boring." │
│ ↳ Model guesses $\rightarrow$ throws in buzzwords and emojis │
│ │
│ Prescriptive Feedback (3-Part Formula): │
│ "Keep paragraphs 1 and 2. In paragraph 3, cut the passive │
│ voice, replace the bullet points with a 3-column table, │
│ and keep the entire response under 100 words." │
│ ↳ 100% Deterministic Correction on Next Turn │
└─────────────────────────────────────────────────────────────┘
The 3-Part Feedback Formula
┌───────────────────────────────────────────────────────────────────────────┐
│ 1. VALIDATE ANCHOR ──► "Keep Section 1 and the introductory hook..." │
│ 2. ISOLATE FLAW ──► "...but Section 2 is too formal and verbose..." │
│ 3. APPLY FIX ──► "...rewrite Section 2 in 2 short punchy sentences." │
└───────────────────────────────────────────────────────────────────────────┘
Master Direct Feedback Prompt Templates
Pattern 1: The Tone & Cadence Calibrator
Your draft is 70% there. Here is the exact feedback:
- **What worked**: The structure and technical points in section 1 are spot-on.
- **What failed**: The tone in the email body is too corporate and apologetic.
- **The Fix**: Rewrite ONLY the email body. Make it confident and conversational, like a senior lead speaking to a peer. Cut out "I hope this finds you well" and "Please let me know if you have any questions".
Pattern 2: The Length & Density Compressor
Good information, but too long.
- Keep the exact same 4 arguments.
- Compress the total length by 50% (ceiling of 150 words).
- Format each argument as: `**Bold Concept**: 1-sentence explanation`.
Pattern 3: The Technical Bug Correction
The logic in function `calculate_tax()` is correct, but your implementation uses an in-memory loop that will cause an OOM error on 100k rows.
- Refactor `calculate_tax()` to use a generator or database SQL aggregation instead.
- Preserve all other functions untouched.
Real-World Case Study
Scenario: Polishing a Proposal Pitch
Ineffective Vague Feedback
User: "This proposal sounds too dry. Make it pop." ❌ AI Output: "🚀 In today's dynamic paradigm, our cutting-edge synergy will unleash unprecedented value! 🔥"
Direct Prescriptive Feedback & Result
User: "The scope breakdown in Section 2 is great. But Section 1 (The Executive Summary) is 3 paragraphs of fluff. Replace Section 1 with a single 3-sentence paragraph stating the current $40k monthly loss and how our solution cuts it to zero within 60 days."
AI Output:
Executive Summary: Your current manual order reconciliation process leaks an average of $40,000 per month in unrecovered billing discrepancies and delays month-end closing by 9 days. Our automated ingestion pipeline eliminates manual reconciliation entirely, reconciling 100% of transactions in real time. We project complete payback within 60 days of deployment.
Summary Best Practices
- Never say "Try again": Always explain what to change and what to keep.
- Give positive constraints: Tell the AI what you want to see ("Use active verbs") rather than just what you dislike.
- Preserve working sections: Explicitly instruct the AI: "Do not modify section X".