3.3 — Instruction-Based QC Agents
Instruction-Based QC Agents: The Final Pass
In 2026, we don't trust the first output of an LLM. We implement Instruction-Based QC (Quality Control) Agents to audit the work of the primary agent before it ever reaches a client.
🏗️ The QC Loop Architecture
- The Creator: Drafts the initial content (e.g., a pitch or a script).
- The Auditor: Compares the draft against a strict technical checklist.
- The Refiner: Rewrites the draft based on the Auditor's feedback.
Technical Snippet: The Auditor System Prompt
### SYSTEM ROLE
You are a Senior Technical Auditor. Your goal is to find 3 reasons to REJECT the provided draft.
### CHECKLIST
1. Does the draft use any 'Forbidden Words' (delve, unlock)?
2. Is the LCP score mentioned correctly?
3. Is the tone 'Institutional Principal' (High Status)?
### OUTPUT
If valid: Return 'PASS'.
If invalid: Return a list of specific 'Fixes' for the Creator.
Nuance: Dual-Model QC
For maximum fidelity, use different models for the Creator and the Auditor. For example, use Gemini 2.5 Pro to create and Claude 4.6 to audit. This prevents "Self-Confirmation Bias" where a model ignores its own mistakes.
Practice Lab: The 3-Step Loop
- Step 1: Ask AI to write a marketing email.
- Step 2: Paste that email into a new thread with the Auditor prompt above.
- Step 3: Feed the Auditor's feedback back to the first thread and ask for a final version.
- Result: Compare the "Pre-QC" and "Post-QC" versions.
📺 Recommended Videos & Resources
-
[AI Quality Control Automation — Multi-Agent Systems] — How to chain creator + auditor agents for production-grade output.
- Type: Video / Tutorial
- Search YouTube for: "AI quality control agents" or "multi-agent orchestration prompts"
-
[The Auditor Checklist Pattern] — Framework for building strict technical QC systems.
- Type: Documentation / Blog Post
- Search: "QC checklist prompts" or "instruction-based review agents"
-
[Dual-Model QC: Claude + Gemini] — Why different models catch different errors (cross-model verification).
- Type: Article / Best Practice Guide
- Search: "multi-model fact-checking" or "dual LLM quality assurance"
-
[Pakistani Agency QC Standards] — Real checklist from Karachi digital agencies on what constitutes "production-ready" content.
- Type: Community Guide / Case Study
- Search for: "Pakistan agency QC standards" or "Karachi content quality checklist"
🎯 Mini-Challenge
5-Minute Task: Run a 3-step QC loop.
Step 1: Creator Agent
"Write a 30-second TikTok script for a Karachi pizza restaurant. Must have a hook in Roman Urdu and 2 scene changes."
Step 2: Auditor Agent
"Review this script. Check:
- Is it under 35 seconds (150 words)?
- Does it have a Roman Urdu hook? ✓/✗
- Are there exactly 2+ scene changes? ✓/✗
- Any forbidden words (delve, unlock, passion)? ✓/✗ If any ✗, list 3 specific fixes."
Step 3: Refiner Agent
"Based on the audit feedback above, rewrite the script to pass all checks."
Challenge: Compare Step 1 (raw) vs. Step 3 (refined). The refined version should be noticeably tighter.
🖼️ Visual Reference
📊 [Instruction-Based QC Loop]
CREATOR AGENT
┌──────────────────┐
│ Drafts Content │
│ (v1.0 - Raw) │
└────────┬─────────┘
│
AUDITOR AGENT
┌────────▼─────────────────────┐
│ Checks against strict rules: │
│ • Forbidden words │
│ • Length constraints │
│ • Tone markers │
│ • Data accuracy │
│ Returns: PASS or LIST OF FIXES│
└────────┬─────────────────────┘
│
┌───────▼─────────┐
│ PASS? │
│ YES────────────────────┐
│ NO │
│ │ │
│ ▼ ▼
│ REFINER AGENT OUTPUT (v2.0)
│ (Fixes issues) Production-Ready
│ │
│ ├──→ AUDITOR (round 2)
│ │
│ └──→ OUTPUT (v3.0) ✓
│
Homework: The Script Auditor
Design a QC agent for a "Faceless Video Script." The agent must ensure the script is under 150 words, contains a Roman Urdu hook, and includes exactly 3 visual "Scene Change" markers.
Lesson Summary
Quiz: Instruction-Based QC Agents: The Final Pass
5 questions to test your understanding. Score 60% or higher to pass.