Stop accepting polite hallucinations. Here are the exact frameworks, templates, and stress-tests to force your AI into brutal self-correction.
You are currently letting a glorified predictive text engine dictate your business strategy, and you're wondering why the results feel a little hollow.
First drafts from an LLM are designed to be agreeable, not accurate… Let me say that again but louder…
FIRST DRAFTS FROM AN LLM ARE DESIGNED TO BE AGREEABLE, NOT ACCURATE!
If you want output that survives contact with the real world, you have to stop asking the AI for answers and start demanding a rigorous self-roast.
🐾 Spro interrupts: "IF IT SMELLS WEAK, WE BITE IT! NO MERCY FOR LAZY TOKENS!"
The secret isn't better instructions. It's structural friction. By forcing the model to switch personas—from a willing creator to a cynical evaluator—you filter out the fluff before it ever reaches your screen.
Here is the complete, unfiltered system for deploying Self-Reflective Evaluation Loops.
1. Full Systems and Workflows
The foundation of this method is a strict, sequential feedback cycle. You cannot let the model generate, evaluate, and revise in the same breath. It must be segmented.
markdown
# System: Core "LLM-as-Judge" Workflow
**Goal:** Generate a high-stakes asset (strategy, code, copy) that has been pre-audited for fatal flaws.
**Step 1: The Raw Generation**
* Prompt: "Analyze the following situation: [Insert Context/Data]. Propose a comprehensive solution with key initiatives, timelines, and expected outcomes."
**Step 2: The Judge**
* Prompt: "STOP. Do not generate further solutions. Switch to strict evaluator mode. You are now a cynical, high-standard auditor. Evaluate your previous output based on:
1. Strategic depth (Is this generic?)
2. Feasibility (Can a small team actually execute this?)
3. Unstated assumptions (What must be true for this to work, but hasn't been proven?)
Score the draft 1-10 and list the top 3 weaknesses."
**Step 3: The Hardened Revision**
* Prompt: "Switch back to the role of the creator. Based *only* on the feedback from the Judge phase, generate a v2 of the solution. Highlight the specific changes made to address the weaknesses."
2. Step-by-Step Templates and Guides
To get elite-level critique, you need to simulate a room full of elite critics. The "Expert Ensemble" template forces the AI to view its work through competing lenses.
markdown
# Template: The "Expert Ensemble" Board of Directors
**Use Case:** Strategic planning, product launches, or complex content creation.
**Prompt to append after your initial draft generation:**
"We are now entering the 'Expert Ensemble' evaluation phase. You will simulate a board of directors critiquing the current draft.
Adopt the following personas one by one and provide a one-paragraph brutal critique from their specific viewpoint:
1. **The CFO:** Focus on budget bloat, ROI timeline, and financial risk.
2. **The Operations Director:** Focus on execution bottlenecks, team bandwidth, and logistical reality.
3. **The Skeptical Customer:** Focus on why they wouldn't buy this, confusing messaging, and lack of trust.
After all three have provided their critique, synthesize their feedback into a 'Mandatory Fixes' list. Finally, rewrite the draft addressing all mandatory fixes."
☕ Mugsy says:"It’s like intentionally inviting the most annoying people you know to a meeting, but you can mute them the second they actually fix your problem."
3. Step-by-Step Breakdowns
For output where failure is expensive, standard critique isn't enough. You need the "Adversarial Stress-Test," designed specifically to simulate external market shocks.
markdown
# Breakdown: Adversarial Stress-Test Workflow
**Goal:** Protect strategies against worst-case scenarios.
**Step 1: Baseline Establishment**
* Execute the standard "LLM-as-Judge" workflow to reach a polished v2 draft.
**Step 2: Injecting the Shock**
* Prompt: "This strategy is currently operating in a vacuum. I am now injecting an external market shock. Assume [Insert Scenario: e.g., 'A major competitor just launched a cloned product with a 50% larger ad budget' OR 'The global supply chain for our core component just stalled for 6 months'].
**Step 3: The Impact Assessment**
* Prompt: "Evaluate the v2 draft against this specific shock. Where does the strategy instantly break? What initiatives become useless?"
**Step 4: The Contingency Branching**
* Prompt: "Produce a final v3 'Hardened' strategy. This version must include a 'Contingency Matrix' detailing exact pivot maneuvers if the market shock occurs."
4. Warnings, Edge Cases, and Optimisation Insights
Do not run these loops blindly. The AI can still fall into specific traps if the criteria aren't sharp enough.
markdown
# Insights: Warnings & Optimisations
* **WARNING - The "Echo Chamber" Effect:** If your evaluation criteria are too vague (e.g., "Make it better"), the AI will just compliment its own writing and change a few adjectives. You must use hard constraints (e.g., "Identify the single most expensive point of failure").
* **EDGE CASE - Context Window Bloat:** Running heavy Generation -> Critique -> Rewrite loops consumes massive token counts. For very long documents, apply the loop to one section at a time rather than the entire 50-page document at once.
* **OPTIMISATION - The "Format Lock":** When asking for the v3 revision, always enforce a strict output format. (e.g., "Output the final v3 as a markdown table with columns for Initiative, Owner, Risk Level, and KPI"). This prevents the model from lapsing back into conversational fluff.
The Insight Reframe: We inherently view friction as a delay. But intentionally manufacturing friction during the AI's drafting process actually speeds up your workflow. You trade 30 seconds of compute time for three hours of manual editing.
The Bottom Line
You now have the complete architecture to stop your AI from lying to you. We've covered:
The core "LLM-as-Judge" sequential workflow.
The "Expert Ensemble" multi-persona template.
The Adversarial Stress-Test for market shocks.
The edge cases and context-window optimisations required to keep it running smoothly.
You can get downloadable, easy-to-import copies of all the Notion/Obsidian Markdown assets above to keep forever right here:
Don't settle for the first draft. Force the system to earn its output. Implement these templates into your prompts today, and watch the quality of your strategic assets multiply.
☕️Until next time - Stay Caffeinated☕️
Thank you for being a paid subscriber.
If this guide helped you - even one prompt, one idea, one shift in how you're thinking about this - the best thing you can do is tell someone.
Recommend CaptionedInCaffeine on Substack to one person who needs it.
That's how we grow. That's how more people find the thing that unsticks them.
Share at captionedincaffeine.substack.com
Stay grounded. Stay curious. And for the love of all things caffeinated - don't let your coffee go cold while you're building something.
©️CaptionedInCaffeine 2026| AI • Productivity • Caffeine Culture • ADHD Lens