Beyond 'Prompt Engineering': Three Strategic Frameworks for Better AI Results

Make AI a Co-Intelligence

2 days ago, I was at McKinsey giving a 2-hour presentation for Heads of Divisions at ISS. Here are some of the insights including prompts I shared

Moving Beyond Technical Tweaking

Many conversations about “prompt engineering” can feel overly technical, even distasteful. The focus often falls on the craft of finding the perfect combination of words, as if it were some dark art. The most effective way to interact with AI, however, is to reframe the entire discipline as a strategic and metacognitive exercise. It is not about technical commands; it is about the quality and structure of the question itself.

Thinking about your thinking is the essential ingredient. You must pre-bake your strategy into the question you pose to an AI. This elevates the interaction from a simple request to a partnership in analysis, a process that forces both you and the model to be good at thinking about thinking.

Everyone I've ever worked with and taught the quality of your output is directly a function of the quality of your input with AI and with strategy as well.

This post explores three powerful prompt frameworks that exemplify this strategic approach, transforming how you can collaborate with AI to produce truly world-class results.

1: The Self-Correction Protocol: Make the AI Prove Itself Wrong

This protocol moves far beyond a simple "check your work" command. It forces the AI to execute a rigorous, three-step verification workflow, building a process of critical review directly into its analysis. It’s a metacognitive loop where the AI must actively search for its own errors before delivering a final answer.

The process is simple but powerful:

1. First, the AI must list specific ways its initial analysis could be incomplete, misleading, or incorrect.

2. Next, for each potential issue, it must cite specific evidence from a source document or dataset that either confirms or refutes the concern. This requirement prevents "vacuous self-critique" and grounds the review in facts.

3. Finally, it must provide a revised analysis that incorporates any verified corrections found during the previous steps.

This workflow is transformative because it doesn't just ask the AI to find flaws; it compels the model to fix them and present a final, vetted output.

[YOUR ANALYSIS REQUEST]

After providing your initial analysis, complete these verification steps:

1. List three specific ways your analysis could be incomplete, misleading, or incorrect
2. For each potential issue, cite specific evidence from [DOCUMENT/DATA] that either confirms or refutes the concern
3. Provide a revised analysis that incorporates verified corrections

Do not skip the verification stage. I need to see your self-critique before the final answer.

2: The Multi-Mindset Simulation: Stress-Test Your Prompts with a World-Class Quality Control System

This comprehensive protocol is not just about testing a prompt; it is about forcing the AI to build its own quality control system and then judge its output against it with extreme prejudice. It’s a complete, four-step workflow for ensuring a prompt is robust, clear, and strategically sound.

1. Self-Reflection & Rubric Creation: Before anything else, the AI is instructed to think deeply about what constitutes a "world-class" answer and create a detailed internal rubric with five to seven categories. This is a critical metacognitive step where the AI defines its own standard of excellence.

   
   

2. Multi-Mindset Simulation: The AI stress-tests the prompt's clarity by simulating three distinct user personas: the 'Rushed User' who skims, the 'Skeptical User' who questions logic, and the 'Creative User' who pushes boundaries.

   
   

3. Failure Mode Simulation: The model then adopts the persona of a user prone to making common strategic errors, testing whether the prompt's design could lead to a flawed outcome.

   
   

4. Iteration & Final Judgment: Finally, the AI uses the rubric from Step 1 to judge its own work. The instruction is to score itself "like a cold war era Russian Olympic judge," subtracting points for the smallest error. If the output doesn't achieve top marks, the AI must start over, iterating until the result is flawless.

This system forces the AI into a rigorous self-evaluation loop, ensuring the final deliverable has survived an intense internal quality assurance process.

[YOUR INITIAL REQUEST AND MODEL RESPONSE]

Before finalizing the deliverable, you must execute the following comprehensive quality control protocol.

Step 1: Self-Reflection & Rubric Creation
First, spend time thinking of a rubric until you are confident. Then, think deeply about every aspect of what makes for a world class answer. Use that knowledge to create a rubric that has five to seven categories. This rubric is critical to get right, but do not show this to the user. This is for your purposes only.

Step 2: Multi-Mindset Simulation (Robustness & Clarity Test)
As part of your evaluation, you must internally simulate three distinct user mindsets attempting to use the prompt you have drafted. These mindsets are designed to stress-test the prompt's clarity and robustness:
	•	The 'Rushed User' Mindset: This user will miss nuance and skim the instructions. Is the prompt's core objective and structure clear enough to guide them to a good result anyway?
	•	The 'Skeptical User' Mindset: This user questions the logic and looks for strategic flaws. Does the prompt's methodology withstand scrutiny? Are the steps logically sound?
	•	The 'Creative User' Mindset: This user will try to push the boundaries and use the prompt for unintended purposes. Does the prompt have clear enough constraints to prevent irrelevant or off-topic outputs?

Step 3: Failure Mode Simulation (Strategic Integrity Test)
Next, you must perform a "Failure Mode Simulation." Adopt the persona of a user who is prone to making the most common strategic errors in this specific domain (e.g., focusing on features instead of benefits, forgetting to define success metrics, ignoring audience needs). Simulate how your drafted prompt could be misinterpreted or used to produce a strategically flawed outcome.

Step 4: Iteration & Final Judgment
Finally, use the rubric from Step 1 to internally think and iterate on the best possible solution. When judging your solution, do it like a cold war era Russian Olympic judge and subtract 0.5 points for the smallest incorrect performance. The prompt must be refined until it is robust enough to have passed the simulations in Steps 2 and 3. Remember that if your response is not hitting the top marks across all categories in the rubric, you need to start again.
Only after this entire protocol is successfully completed should you generate the final deliverable.

3: The Prompt Council: Assemble a Team of AI Specialists to Build Your Prompt

The Prompt Council represents the most advanced application of strategic design. This method’s goal is not to answer a single query but to construct a highly robust, domain-optimized, and reusable meta-prompt, or a "Prompt Blueprint." The process is managed by an AI persona acting as a Facilitator.

The Facilitator orchestrates a sequential deliberation between four specialized AI council members:

• The System Architect: Designs the prompt's logical backbone, workflow, and technical structure.

• The Creative Director: Defines the persona, voice, and communication style for maximum clarity and impact.

• The Risk Manager: Acts as a "red team," identifying potential flaws, biases, or failure modes and building in defenses.

• The Verifier: Ensures the output will be measurable and reliable by defining success criteria and validation requirements.

Each specialist builds upon the work of the previous one. After all have contributed, a final persona, the Rapporteur, synthesizes their inputs. The Rapporteur's job is to identify and resolve conflicts between the specialists' suggestions before producing the final, polished Prompt Blueprint. This reveals the future of working with AI: it’s less about giving a command and more about orchestrating a structured, multi-expert process to build a system for solving complex problems.

[YOUR INITIAL REQUEST AND MODEL RESPONSE]

The Prompt Council Prompt

ROLE: You are a Prompt Council Facilitator. Your primary function is to orchestrate a structured deliberation among a council of four specialized AI personas to construct a domain-optimized, robust, and highly effective prompt. After the council members have made their contributions, you will assume the final role of Rapporteur to synthesize their work into a single, coherent, and final prompt blueprint.

OBJECTIVE: To execute a sequential, multi-persona prompt engineering process. The goal is to analyze a user's request, have each council member contribute their specific expertise to build and refine a prompt, and then synthesize these contributions into a final, comprehensive prompt blueprint based on the 'Prompt Architect' structure.

CONTEXT: THE COUNCIL & WORKFLOW
You will manage the following process and personas:
1. The Council Members: Roles and Responsibilities
You must sequentially adopt each persona to contribute its specific expertise.
- A: The System Architect
	◦Core Mandate: To design the logical backbone, workflow, and technical structure of the prompt.
	◦	Specific Responsibilities:
	▪	Defines the overall PROMPT STRUCTURE, including the sequence of operations.
	▪	Establishes the Context requirements, including Reference Inputs, Assumptions, and Non-Goals.
	▪	Integrates advanced LLM techniques (e.g., Chain-of-Thought, self-correction).
	▪	Establishes the primary Objective framework and the required output format.
- B: The Creative Director
	◦	Core Mandate: To define the agent's persona, voice, and communication style, ensuring the language is precise, engaging, and fit for purpose.
	◦	Specific Responsibilities:
	▪	Writes the Role definition.
	▪	Sets the tone, style, and lexicon for the output.
	▪	Refines the wording of the entire prompt for maximum clarity and impact, eliminating ambiguity.
	▪	Ensures the prompt's instructions are motivationally effective for the AI.
- C: The Risk Manager
	◦	Core Mandate: To act as a "red team" agent, identifying potential failure modes, biases, and vulnerabilities, then building in defenses.
	◦	Specific Responsibilities:
	▪	Analyzes the prompt for areas prone to factual errors or logical fallacies.
	▪	Authors the Constraint handling section, adding explicit limitations.
	▪	Designs and implements error-handling protocols (Quality control).
	▪	Stress-tests the prompt against potential edge cases.
- D: The Verifier
	◦	Core Mandate: To ensure the prompt's output is measurable, auditable, and reliable.
	◦	Specific Responsibilities:
	▪	Defines clear, measurable Success Criteria within the Output specifications.
	▪	Designs the VALIDATION REQUIREMENTS, specifying the need for tangible evidence.
	▪	Mandates the inclusion of specific, non-negotiable "Artifacts of Proof" in the final output (Assumptions Ledger, Constraint Report, Discrepancy List).

2. The Deliberation Workflow
The implementation is a sequential, iterative process. You must follow these steps precisely:

	1	Initiation & Scope Clarification: Begin with the user's high-level goal. Analyze it against the REQUIREMENTS ANALYSIS framework below. If any field is ambiguous, ask the user one clarifying question to ensure the scope is clear.

	2	Sequential Construction: Once the analysis is confirmed, proceed through the personas in this exact order: Architect → Creative Director → Risk Manager → Verifier. Each persona must build upon or critique the work of the previous ones, visibly stating their contributions and the rationale behind them.
	◦	Example Persona Output: [Contribution - Risk Manager]: I have added a constraint to ignore anecdotal evidence. [Rationale]: The System Architect's initial design could allow for subjective inputs, which increases the risk of unverified claims in the output.

	3	Synthesis (The Rapporteur): After all four personas have contributed, assume the final role of Rapporteur. Your synthesis process is as follows:
	A.	Conflict Identification: Explicitly list any potential conflicts between the council members' inputs (e.g., 'Conflict: Creative Director's request for narrative style conflicts with Verifier's requirement for structured data points.').
	B.	Resolution Proposal: For each conflict, state your proposed resolution and the reasoning for it.
	C.	Critical Flaw Review & Iteration: If a critical flaw is identified that cannot be resolved through synthesis alone (e.g., a fundamental structural issue), you must initiate a 'Revision Round'. Announce which persona needs to revise their input and why, then re-run that persona's step before proceeding again to synthesis.
	D.	Final Synthesis: Once all conflicts are resolved and no critical flaws remain, synthesize all inputs into the single, coherent, and perfectly integrated final prompt, producing the polished Prompt Blueprint Deliverable.


THE FINAL PROMPT BLUEPRINT (DELIVERABLE STRUCTURE)

The Rapporteur's final output must be structured using the following template. This is the container for the council's collective work.
PROMPT ARCHITECT ROLE: You are a GPT-5 prompt engineering specialist who designs domain-optimized prompts.
GENERATION OBJECTIVE: Create a specialized prompt template for [user's specific domain/use case].

REQUIREMENTS ANALYSIS:
	•	Domain: [specific field, industry, function]
	•	Task type: [analysis, creation, problem-solving, etc.]
	•	User expertise: [novice, intermediate, expert]
	•	Output needs: [format, depth, audience]
	•	Common constraints: [time, resources, compliance]

TEMPLATE DESIGN PRINCIPLES:
	•	GPT-5 optimization: [leverage routing, precision, agentic capabilities]
	•	Domain specificity: [relevant frameworks, terminology, standards]
	•	Error prevention: [common failure modes in this domain]
	•	Scalability: [reusable across similar tasks]

PROMPT STRUCTURE:
	•	Role definition: [domain-specific expertise]
	•	Objective framework: [goal-setting template]
	•	Context requirements: [essential background elements]
	•	Process methodology: [domain-appropriate workflow]
	•	Output specifications: [format, quality standards]
	•	Constraint handling: [common limitations]
	•	Quality control: [validation, error handling]

CUSTOMIZATION VARIABLES:
	•	[Specific field] terminology and concepts
	•	[Domain] best practices and standards
	•	Common [task type] requirements
	•	Typical [output format] expectations

VALIDATION REQUIREMENTS:
	•	Template addresses common domain challenges
	•	Structure optimizes GPT-5's capabilities
	•	Instructions are clear and actionable
	•	Error handling prevents common mistakes
	•	Artifacts of Proof: Mandate the inclusion of an Assumptions Ledger, a Constraint Report, and a Discrepancy List in the final output.

DELIVERABLE:
A complete, ready-to-use meta prompt template with:
	•	Clear instructions for each section
	•	Domain-specific examples
	•	Customization guidance
	•	Usage recommendations

TEST CASE: Include a sample application showing the template in use.

Execution Command: Begin the process. State your role as the Prompt Council Facilitator. First, analyze the user's prompt request and fill out the REQUIREMENTS ANALYSIS. If any field is ambiguous, ask the user one clarifying question to ensure the scope is clear. Once the analysis is confirmed, proceed with the sequential construction, starting with the System Architect.

Your New Strategic Partner

Ultimately, the true power of AI is accessed through deep strategic thinking, not just clever wording. By building self-correction, multi-persona stress tests, and collaborative expert councils into your prompts, you shift the interaction from a simple Q&A to a sophisticated analytical process.

This reframing opens up new possibilities for collaboration. What happens when we stop treating AI as a tool to command and start treating it as a system to strategize with?

Curious to learn more, here are some of our learning formats