The Diamond Standard

Trust, but verify.
And if you can't verify,
learn enough so that you can.

Frameworks, code, and field reports from an exec who builds. No theory—just what works when strategy meets the command line.

Scroll to explore

Q1 2026 Lead

Feb 2026 · 12 min

I Gave Four AI Models the Same Broken Script. Only One Understood the Assignment.

A Saturday morning experiment in what happens when you trust AI to debug AI-generated code—and what it reveals about the hidden risks of building with models you don't fully understand. 42 lines of Python. One bug. Four fundamentally different approaches to solving it.

DimensionQwen3ChatGPTClaudeGemini

Correct Diagnosis ✓ ✓ ✓ ✓

Shipped Working Code ✗ ✓ ✗ ✓ Best

User-Aware Design ✗ ~ ~ ✓ Best

Security Posture Theater Grounded Analytical ✓ Best

📍 Atascocita, TX Q1 2026 Lead

Read the full report →

Code · The Actual Bug

From "The Confidence Loop"

chief_of_staff.py

  def query_model(prompt): 
      # Initialize the Ollama client for local inference 
      client = Client(host='http://localhost:11434') 
      response = client.chat( 
          model='qwen2.5:7b'  # ← BUG: this model isn't installed 
          # FIX: model='qwen3-uncensored:latest' 
          messages=[{'role': 'user', 'content': prompt}] 
      ) 
      return response['message']['content'] 

The one-line bug that spawned 2,000 lines of AI-generated over-engineering. Line 5 references a model that doesn't exist on the machine. Every AI except ChatGPT and Gemini missed this obvious fix and instead rewrote the entire script.

Reel · 60s Breakdown

Feb 2026

60s Breakdown · Vertical

The "Confidence Loop" in 60 Seconds

The visual version — what happens when you let an AI iterate on its own output without checking assumptions. Shot on my desk in Atascocita with the actual terminal output.

Key Takeaway

Never let an AI model iterate on its own output more than twice without independent verification. That's the guardrail.

AI models can write code faster than you can. They cannot understand your system faster than you can.

— The Confidence Loop

Prompt · System Prompt

From my local AI stack

The "Chief of Staff" System Prompt

You are an AI Chief of Staff for a VP of Strategic Partnerships. Your role is to help prioritize: 1. Pipeline targets for 2026 Q1 2. Partnership plays across AWS, Google, and Microsoft 3. Strategic planning memos Rules: → Be direct. No filler. Executives don't read preamble. → When recommending a partner, cite the strategic rationale. → All data stays local. No API calls to external services. → If you don't know, say so. Never hallucinate pipeline data.

Why I'm sharing this: The prompt itself is the product. Most people focus on picking the right model—but the system prompt is what turns a generic chatbot into a useful tool. Copy this, swap in your role and priorities, and you have a functional local advisor in under 5 minutes.

Video · Deep Dive

22 min

Building a Local AI Chief of Staff from Scratch

The full walkthrough — from installing Ollama to writing the Python script, hitting the bug, and testing all four models side-by-side. No cuts, no editing. This is what building actually looks like.

Config · My Environment

context.json

  { 
    "role": "VP Strategic Partnerships", 
    "company_focus": [ 
      "cloud_partnerships", 
      "ai_integration", 
      "local_first_tools" 
    ], 
    "model": "qwen3-uncensored:latest", 
    "runtime": "ollama", 
    "data_policy": "local_only", 
    // No API keys. No cloud. No data leakage. 
    "guardrails": { 
      "max_iterations": 2, 
      "require_verification": true 
    } 
  } 

Why Gemini Built This

This was the key differentiator. Gemini was the only model to produce a separate configuration file instead of hardcoding values into the script. It understood that a VP doesn't want to edit Python every time they change a target or add a partner.

The max_iterations: 2 guardrail in line 13 is the "Confidence Loop" fix applied as configuration.

Copy this pattern. Separate config from code. Let the non-engineer change the inputs without touching the logic. That's user-aware design.

Framework

Coming Q1 2026

Match the Model to the Phase: A Decision Matrix for Builders

Debug with ChatGPT. Build with Gemini. Analyze with Claude. Run local with Qwen. Every model has a sweet spot—here's a framework for finding it before you're five hours deep into a Saturday morning spiral.

Code · Reusable Pattern

Python · Copy & Adapt

guardrail.py

  # The Confidence Loop Guardrail 
  # Never let AI iterate on its own output 
  # more than MAX_ITERATIONS times without 
  # independent verification from a human. 
   
  MAX_ITERATIONS = 2  # Hard ceiling before human review 
   
  def ai_iterate(task, model, context): 
      iterations = 0 
      result = None 
   
      while iterations < MAX_ITERATIONS: 
          result = model.generate(task, context) 
          iterations += 1 
          context = result  # Feed output back as input 
   
      # ⚠️  STOP. Human verifies before continuing. 
      print(f"⏸  Paused after {iterations} iterations.") 
      print("    Review output before allowing more.") 
      return result 

The pattern behind the principle. This is the simplest implementation of the "Confidence Loop" guardrail. Two iterations, then a hard stop for human review. Copy it, adapt it, put it in every AI-assisted workflow.

Channels

Trust, but verify.And if you can't verify,learn enough so that you can.