Leadership Guide 2026: The Art of Having a Conversation With AI

From 'System 1' to 'System 2' dialogue with AI

10 techniques for better AI answers

In late November and December 2025, new versions of the major AI models were released. With that came a new generation of language models that can reason in new ways. They can hold more requirements in mind at once, they can genuinely assess their own answer, and they can correct their answer before you see it, without you having to spend many rounds pushing them in the right direction.

It means that in leadership work you can now get finished answers and outcomes in the first output, and you can use them directly for decisions, communication, or plans, because you can see what the answer is based on.

It also means that techniques such as Chain of Thought are outdated. OpenAI describes directly how the new models do this themselves, and they recommend that you stop using the technique. You get more value from describing your goal, your requirements, and your constraints clearly, then demanding an answer that shows uncertainties, risks, and what is missing before you act.

This leadership guide gives you 10 techniques for that. These are not random techniques. They are ones shared at different times over the last month by the people who built the models.

I have rewritten them for leadership tasks. Primarily for myself, and for the situations where I need a result I can trust and therefore pass on.

The techniques help you make the model’s work visible enough that you can assess quality quickly. Each technique has a clear template, a filled example, and what it enables you to do, so you can adapt it into your own context quickly.

Use the guide in ChatGPT 5.2, Gemini 3, Claude 4.5, and Microsoft 365 Copilot. Choose two or three techniques that fit your tasks, and use them consistently for a couple of weeks. When you want your model to write a good prompt or a good question and your two or three techniques do not match the task, attach the guide and ask the model to choose the technique that fits best and fill the template with your context. That makes your AI answers noticeably more stable, and it significantly reduces the time from first keystroke to finished result.

1. Chain of Verification

Google’s research team uses this technique to eliminate hallucinations.

Template

Task: [Insert your request]

Step 1: Provide your initial answer.

Step 2: Create 5 verification questions that can reveal errors or weak assumptions in your answer.

Step 3: Answer each verification question. Mark what is documented knowledge and what is assumptions.

Step 4: Deliver a final, corrected answer based on the verification.

Example

Task: I need to recommend a prioritisation of three initiatives for next quarter, and I need to send a decision memo to the leadership team.

Step 1: Provide your initial answer.

Step 2: Create 5 verification questions that can reveal errors or weak assumptions in your answer.

Step 3: Answer each verification question. Mark what is documented knowledge and what is assumptions.

Step 4: Deliver a final, corrected answer based on the verification.

Effect

You get an answer built as a process with checks built in, so errors and weak assumptions become visible before they turn into a memo or a recommendation. You get a clear marking of what is knowledge and what is assumptions, so you can assess risk and ask the right follow up questions. You become less dependent on the first draft and more dependent on what has been verified, which gives a stronger basis to pass on to the leadership team or the board.

2. Meta-Prompting

This is what OpenAI’s red team uses to break their own models and find edge cases.

Template:

I need to achieve: [overall goal]

Your task:

1. Analyse what would make a very strong prompt for this goal.

2. Consider: specificity, context, constraints, output format, and the examples that are needed.

3. Write the prompt.

4. Run the prompt and deliver the output.

[GOAL]: [your actual goal]

Example:

I need to achieve: A clear brief for a leadership meeting about a new organisational priority, ready to send without rewriting.

Your task:

1. Analyse what would make a very strong prompt for a meeting brief for leaders that needs to be both short and decision ready.

2. Consider: audience, context, constraints, output format, the examples that are needed, and the misunderstandings that typically occur.

3. Write the final prompt.

4. Run it and deliver the result.

GOAL: One page meeting brief with a clear decision, 3 expected objections with responses, and a closing recommendation on next steps.

Effect

You get a precise task specification that fits your situation, your requirements, and your audience before the model starts writing the actual text. You get a result closer to send ready quality on the first attempt because quality is governed by requirements and format, insted of by repeated corrections. You also make visible where the task is typically misunderstood, so you can close gaps early and save time on follow up explanations.

3. Role-Based Constraints

The technique defines expertise and boundaries for the task in advance.

Template:

• Role: You are a [specific role] with [X years] of experience in [domain].

• Task: [Specific task].

• Constraints: [List 3 to 5 specific constraints].

• Output format: [Precise format].

Example

You are an organisational consultant with 12 years of experience in meeting practice in Danish B2B companies.

Your task: Write a decision memo for the leadership team on reducing meeting time by 20 percent over the next 8 weeks without losing decision capacity.

Constraints: Use at most 220 words. Use short sentences. State 3 concrete levers. State 2 risks with mitigation. End with a decision recommendation and a next step with an owner.

Output format: Title. Context in 3 lines. Decision. Rationale. Levers. Risks. Next step.

Effect

You get an answer that hits the professional angle, the tone, and the form you need because the role, boundaries, and format are set before the model writes. You get output that can be forwarded with minimal editing because length, structure, decision requirements, and risks are built into the request. You get a pattern that can be reused in your team, so quality stays stable even when multiple people write and the pace is high.

4. Structured Thinking Protocol

OpenAI’s GPT 5 team uses this approach for complex reasoning tasks.

Template

Complete these steps before the final answer:

[UNDERSTAND]: Restate the problem and identify the actual question.

[ANALYZE]: Break it into components. Note assumptions and constraints.

[STRATEGY]: Sketch 2 to 3 possible approaches and assess trade offs.

[EXECUTE]: Provide the final answer with rationale.

Question: [Your question]

Example

Before you answer, complete these steps:

[UNDERSTAND]

• I need to decide whether we should centralise or decentralise a support function.

• I am being asked for a recommendation, consequences, and next steps.

[ANALYZE]

• Components: economics, quality, speed, accountability, culture.

• Assumptions: current bottlenecks, competence level, system support.

[STRATEGY]

• Approach 1: Centralise for consistency and control.

• Approach 2: Decentralise for proximity and speed.

• Trade offs to be decided by the leadership team: governance, responsiveness, cost, risks.

  
  

[EXECUTE]

• Recommendation: Choose centralisation with local contact points.

• Criteria: decision time, consistent service, clear accountability.

• Next step: pilot in 2 teams for 6 weeks with clear measures.

Question: What is the most decision ready way to implement the change, and which compromises do we accept.

Effect

You get an answer that starts with a clear understanding of the problem and ends with a recommendation where assumptions, constraints, and trade offs are visible. You get multiple possible paths forward with clear consequences, which makes it easy to choose what to test, what to decide, and what needs more data. Disagreement in leadership becomes easier to handle because it can be placed on criteria and consequences, and because next steps can be formulated with ownership and timing.

5. Few-Shot with Negative Examples

Anthropic has shown that explicit negative examples strengthen the model’s understanding of quality.

Template

I need [task]. Here are examples:

GOOD example: [Insert text that meets your standard]

BAD example: [Insert unwanted text]

Reason: [Explain why the example is bad]

Now write: [Your actual task]

Example

I need you to write an internal email about a changed priority. Here are examples:

GOOD example 1: Short. Decision first. Consequence stated. Next step clear.

GOOD example 2: Clear owner. No clichés. Respect for the reader’s time.

BAD example 1: Long introduction about why the world is changing.Why it is bad: It avoids the decision and creates uncertainty.

BAD example 2: Vague wording with no owner or timeline.Why it is bad: No one can act on it.

Now write: An email of 120 to 160 words to all employees stating that we are temporarily prioritising operational stability over new initiatives for the next 4 weeks. The email must include the decision, why, what stops, what continues, and what I expect from line managers.

Effect

You teach the model your quality standard by showing what is good, what is unacceptable, and why. You get output that follows your tone, length, and decision requirements with fewer corrections because the standard is fixed before it writes. You can reuse your examples across emails, memos, and briefs, so quality stays stable even when you work fast.

6. Confidence-Weighted Prompting

Google DeepMind applies this method for high risk decisions.

Template

Answer the following question: [Question]

In your answer, include:

1. The primary answer with a recommendation.

2. Confidence level as a percentage (0 to 100%).

3. Key assumptions behind the answer.

4. What would change the answer.

5. An alternative answer if the confidence level is below 80%.

Example

Answer this question: Should we stop an ongoing project that is delayed.

In your answer, include:

1. Your primary answer with a clear recommendation.

2. Confidence level as a percentage from 0 to 100.

3. Key assumptions, especially about remaining work, dependencies, and alternative use of capacity.

4. What would change your answer and what data you are missing.

5. An alternative answer if the confidence level is below 80 percent, and what you would do first.

Effect

You get a recommendation with visible confidence, key assumptions, and a clear description of what could change the answer, so the decision can be tied to the next clarification. You reduce the risk that a confidently worded answer is treated as knowledge when it actually rests on assumptions. You get an alternative track when confidence is low, so you can still act, but within a clear risk frame.

7. Context Injection with Boundaries

Anthropic’s prompt engineers use extensive context to ensure traceability.

Template

[CONTEXT]: [Insert your documentation or report]

[FOCUS]: Use only information from CONTEXT. If data is missing, answer: "Insufficient information".

[TASK]: [Your specific question].

[CONSTRAINTS]: Refer to specific sections in CONTEXT. Do not use knowledge outside the material.

Example

[CONTEXT]

Draft meeting policy:

1. Meetings are scheduled only when the purpose is a decision or a concrete clarification.

2. All meetings have an owner, an agenda, and an expected output.

3. Status updates move to written updates no later than the day before.

4. Standard meeting length is 25 minutes. Exceptions require justification.

5. Decisions are documented the same day with an owner and a deadline.

[FOCUS]

Use only information from CONTEXT to answer. If the answer requires something that is not in the text, state what is missing.

[TASK]

Rewrite the meeting policy so it is more decision oriented, clear on accountability, and ready to be sent across the organisation as a short standard.

[CONSTRAINTS]

• Point out which sentences you changed and why.

• Do not add general advice outside the text.

• If multiple interpretations exist, mention them and choose one with a rationale.

Effect

You get a revised document that is built on your material, and an explicit marking of what is missing in the text for the answer to become more precise. You get traceability to specific places in your basis, so you can see what carries the changes. That makes the output suitable for internal documents because the model stays within your context and avoids filling gaps with generic assumptions.

8. Iterative Refinement Loop

OpenAI’s research team uses this sequence to optimise quality through critique.

Template

[STEP 1]: Draft [task].

[STEP 2]: Review the draft. Identify 3 weaknesses or gaps.

[STEP 3]: Rewrite the draft so the identified weaknesses are addressed.

[STEP 4]: Is this ready to send? If not, what is missing?

Example

[Step 1]

Draft a one page plan for onboarding new leaders in my department.

[Step 2]

Review the draft. Identify 3 weaknesses, especially unclear ownership, unclear goals, and missing repeated practice.

[Step 3]

Rewrite the plan so the weaknesses are resolved. Keep it to one page.

[Step 4]

Final review: Is this ready to be sent. If not, specify exactly what is missing and suggest corrections.

Effect

You get a draft, a concrete critique, and an improved version in one fixed sequence, so quality increases without you having to write each iteration. You get a quality gate where the model checks send readiness and points precisely to gaps, so you can stop early when something does not hold. It reduces your editing time and makes it easier to standardise what is allowed to be forwarded.

9. Constraint-First Prompting

Google Brain researchers prioritise the constraints before the task itself to minimise errors.

Template

Hard constraints (non negotiable):

[Constraint 1, for example: Max 180 words]

[Constraint 2, for example: Direct leadership style suited for Danish workplaces]

[Constraint 3, for example: Must end with a clear decision]

Soft preferences (desirable):

[Preference 1, for example: Calm and precise tone]

Task: [Your request] Repeat the constraints in one sentence before you answer.

Example

Hard constraints (cannot be violated):

• Write in English in a concise leadership style without clichés.

• Use at most 180 words.

• End with a clear decision and a next step with an owner.

Soft preferences (optimise for these):

• The tone should be calm and precise.

• Use a few concrete examples.

Task: Write a short brief to my team that we are changing priorities for the next 30 days, and that we are temporarily stopping two side initiatives.

Repeat the constraints in one sentence before you answer.

Effect

You get a consolidated recommendation where trade offs are visible in advance, and where multiple legitimate considerations are handled concretely before the decision. You get a shared language for handling disagreement because the perspectives make it clear what is being assessed and what the consequences are. You get a suggestion for a pilot or test that reduces risk and makes the decision robust to challenge.

10. Multi-Perspective Prompting

Anthropic’s Constitutional AI uses multiple viewpoints to increase robustness.

Template

Analyse [issue] from these perspectives:

[PERSPECTIVE 1]: Operational feasibility.

[PERSPECTIVE 2]: Business impact.

[PERSPECTIVE 3]: Customer and employee experience.

[PERSPECTIVE 4]: Operability and accountability.

Synthesis: Integrate all perspectives into a final recommendation with clear trade offs.

Example

Analyse the decision to change customer service opening hours from these perspectives:

[PERSPECTIVE 1: Technical feasibility]

Assess the systems, processes, and dependencies that are affected, and what needs to change to deliver reliably.

[PERSPECTIVE 2: Business impact]

Assess the impact on cost, service level, sales, and customer loyalty. State your assumptions.

[PERSPECTIVE 3: User experience]

Assess the consequences for customers and for customer service employees, including friction and expected resistance.

[PERSPECTIVE 4: Risk and compliance]

Assess risks, including reputation, contractual obligations, and what could go wrong in implementation.

Effect

You make legitimate considerations visible before the decision, which makes it robust to challenge. You get a consolidated recommendation with clear trade offs, so the dialogue stays focused on criteria and consequences.