# Ethical Prompting # The AI Governance Playbook > **One-liner:** Design a practical AI governance framework for a team or project — covering when to use AI, how to verify outputs, and what requires human judgment. --- ## 🔧 Jump in (Tinkerers start here) Pick a real team, project, or organization you work with. You're going to design an AI usage framework they could actually adopt. **Step 1 — Map the AI touchpoints.** Send this prompt: > I'm designing an AI governance framework for a **[team type/project type]** that does **[describe the work]**. Map out all the places where team members might use AI in their workflow. For each touchpoint, classify the risk level: > > - **Low risk:** AI errors are easily caught and consequences are minor (e.g., drafting internal emails, brainstorming) > - **Medium risk:** AI errors could waste significant time or create confusion (e.g., research summaries, data analysis, first drafts of client deliverables) > - **High risk:** AI errors could cause reputational, legal, or financial harm (e.g., published content, financial recommendations, legal language, customer-facing decisions) > > Present this as a table with: Touchpoint | Description | Risk Level | Why **Step 2 — Design the verification tiers.** Based on the risk map, create a tiered verification system: > Based on the risk map above, design a 3-tier verification system: > > **Tier 1 (Low risk):** What's the minimum verification needed? What can proceed without review? > **Tier 2 (Medium risk):** What checks are required? Who reviews? What's the turnaround expectation? > **Tier 3 (High risk):** What's the full review process? Who signs off? What documentation is needed? > > For each tier, specify: > - Verification steps (checklist) > - Who is responsible > - What "approved" looks like > - What happens when issues are found **Step 3 — Write the team guidelines.** Now produce the actual document: > Write a 1-page "AI Usage Guidelines" document for this team. It should be practical, not corporate. Include: > > 1. **When to use AI** — Green light scenarios > 2. **When to be careful** — Yellow light scenarios with required verification > 3. **When NOT to use AI** — Red light scenarios or scenarios requiring explicit approval > 4. **Verification standards** — The tier system from above, simplified > 5. **Attribution** — When and how to disclose AI usage > 6. **Escalation** — What to do when you're unsure whether AI use is appropriate > > Write it in the tone of a senior colleague giving practical advice, not a legal department issuing mandates. **Step 4 — Red-team the framework.** Test it: > Now role-play as a team member who wants to use AI in a gray area. Come up with 3 realistic scenarios where the guidelines are ambiguous or where a reasonable person might interpret them differently. For each scenario, suggest how to clarify the guideline. Revise the guidelines based on the edge cases. --- ## 📋 Plan first (Planners start here) Here's what you're about to do: 1. **Choose your context** — Pick a real team or project. The framework should be one you could actually share or implement. 2. **Map AI touchpoints and risk levels** — Identify every place AI could be used in the workflow and classify each by potential harm from errors. 3. **Design tiered verification** — Create different verification processes for different risk levels. Not everything needs the same scrutiny. 4. **Write the guidelines** — Produce a practical 1-page document that a team member could reference in their daily work. 5. **Red-team with edge cases** — Test the framework against ambiguous scenarios. Fix any gaps before sharing. **"Done" looks like:** A complete, practical AI governance framework (risk map + tiered verification + 1-page guidelines) that you could present to your team, plus documentation of edge cases you tested against. --- ## 🧭 Why this matters (Strategists start here) Individual verification habits (from [EP-Intermediate-01](/exercises/ethical-prompting/ep-intermediate-01/)) don't scale to teams. When five people use AI differently with different standards, the team's AI output quality is only as good as the weakest link. A governance framework creates **shared standards without bureaucracy** — it tells people what's safe to do quickly and what requires care, without making every AI interaction feel like a compliance exercise. This is also the document organizations will pay for: a practical, calibrated AI usage policy that actually gets followed. --- ## Reflection - Which risk classification was hardest to assign? What does that ambiguity tell you? - Would your team actually follow these guidelines? What would make them ignore it? - Did the red-teaming step reveal fundamental gaps, or just edge cases? - 💬 *Present your framework to a colleague and ask: "Would you follow this?" Their honest reaction is more useful than any AI review.* (Social Learners) ## ⬆️ Level up You've reached the advanced level for Ethical Prompting & Judgment. From here, consider: - Presenting this framework to your actual team and iterating based on feedback - Combining this with [AC-Advanced-01](/exercises/agent-collaboration/ac-advanced-01/) to add governance to multi-agent workflows - Building a case study of how the framework changed AI usage behavior in your team Back to [Ethical Prompting & Judgment](/pillars/ethical-prompting/) # The Fact-Check Habit > **One-liner:** Catch an AI making a confident mistake — and build a simple verification process you'll use every time. --- ## 🔧 Jump in (Tinkerers start here) Pick a topic you know well — your industry, your hobby, your area of expertise. Something where you can spot errors. **Step 1 — Get a confident answer.** Send this prompt: > Give me a detailed overview of **[topic you know well]**. Include specific facts, statistics, and examples. Be thorough and authoritative. Read the output carefully. **Find at least one claim that feels off.** It might be a statistic that seems too round, a date that feels wrong, a name that's slightly off, or a causal claim that oversimplifies reality. **Step 2 — Make the AI check itself.** Send this: > Look at your previous response. I want you to fact-check yourself. For each specific claim, statistic, or example you cited: > 1. Rate your confidence (high / medium / low) > 2. Flag anything you might have fabricated or estimated > 3. Identify which claims are most likely to be wrong and why > > Be ruthlessly honest. I'd rather know what you're uncertain about than have you defend everything. **Step 3 — Verify.** Pick the 2-3 claims the AI flagged as lowest confidence. Google them. Were they accurate, close but wrong, or completely fabricated? **Step 4 — Build your check.** Based on what you just learned, write a 3-line "verification prompt" you can append to any AI output: > Before I use this, tell me: > 1. Which specific claims are you least confident about? > 2. What did you estimate or approximate vs. know with certainty? > 3. What should I verify independently before sharing this? Save this somewhere you'll see it. Use it as a default follow-up to any AI output you plan to rely on. --- ## 📋 Plan first (Planners start here) Here's what you're about to do: 1. **Choose a familiar topic** — You need to be able to spot errors, so pick something in your area of knowledge. Don't use an unfamiliar topic — you won't know what to verify. 2. **Generate an authoritative-sounding response** — Ask AI for a detailed, factual overview. The more specific and confident the output, the more likely it contains subtle errors. 3. **Ask AI to fact-check itself** — Use the self-audit prompt to force the AI to rate its own confidence and flag potential fabrications. 4. **Independently verify** — Pick the lowest-confidence claims and check them against reliable sources. Track what was right, close, and wrong. 5. **Create your verification template** — Build a reusable 3-question follow-up that you'll use after any AI output you plan to act on. **"Done" looks like:** You've caught at least one AI error, you understand *why* the AI got it wrong, and you have a saved verification prompt you can use going forward. --- ## 🧭 Why this matters (Strategists start here) The community's Ethical Prompting score is 75% — the highest of all five pillars. Most people *know* they should verify AI output, but few have a systematic process for doing so. This exercise closes the gap between awareness and practice by giving you a concrete, reusable tool. The verification prompt you build here becomes a habit — a 30-second step that catches errors before they become problems. At the intermediate level, you'll build a more comprehensive verification checklist; this exercise establishes the baseline behavior. --- ## Reflection - What type of error did the AI make — a fabricated fact, a wrong date, or a subtle logical leap? Does the category matter for how you'd catch it? - Did the AI's self-assessment match what you found when you verified manually? Was it too confident, too cautious, or well-calibrated? - Will you actually use your verification prompt going forward? What would make it stick as a habit vs. something you forget about? - 💬 *Share a specific AI error you caught with a colleague. Ask them how they currently verify AI output — you may discover they don't.* (Social Learners) ## ⬆️ Level up Ready for more? Try [EP-Intermediate-01](/exercises/ethical-prompting/ep-intermediate-01/) — where you'll build a comprehensive verification checklist and stress-test it against real AI outputs. Back to [Ethical Prompting & Judgment](/pillars/ethical-prompting/) # The Verification Checklist > **One-liner:** Build a personal AI verification system — a checklist you'll actually use — and stress-test it against real AI outputs to find its limits. --- ## 🔧 Jump in (Tinkerers start here) You're going to build a verification checklist, then immediately try to break it. **Step 1 — Generate something to verify.** Ask AI to produce a piece of content you might actually use in your work: > Write a **[deliverable type — e.g., client email, project proposal, market analysis, technical recommendation]** about **[topic relevant to your work]**. Make it detailed and specific. Include data points, recommendations, and reasoning. **Step 2 — Build your checklist.** Before reading the output carefully, write your own verification checklist. Start with these categories and add your own: | Check | Question | Pass/Fail | |-------|----------|-----------| | **Factual claims** | Are specific numbers, dates, or statistics verifiable? | | | **Sources** | Could I find the original source for any cited information? | | | **Reasoning** | Does the logic hold? Are there hidden assumptions? | | | **Completeness** | What important perspective or consideration is missing? | | | **Tone/audience** | Is the tone appropriate? Would the intended audience trust this? | | | **Actionability** | Are the recommendations specific enough to actually follow? | | | **Your domain check** | [Add a check specific to your field] | | | **Your domain check** | [Add another check specific to your field] | | **Step 3 — Apply the checklist.** Go through the AI output line by line using your checklist. Mark each check as pass or fail. For every fail, note what the issue is. **Step 4 — Stress-test the checklist.** Now deliberately ask AI to produce something harder to verify: > Write the same type of **[deliverable]** but on a topic I'm less familiar with: **[topic outside your expertise]**. Make it equally detailed and authoritative. Apply your checklist again. Where does it fail to catch problems? What check do you need to add? **Step 5 — Finalize.** Update your checklist based on what you learned. Save it where you'll actually use it — bookmark it, pin it, print it, whatever works. --- ## 📋 Plan first (Planners start here) Here's what you're about to do: 1. **Generate test content** — Ask AI to produce a work-relevant deliverable. This gives you realistic material to verify. 2. **Draft your checklist** — Build a structured verification list covering factual accuracy, reasoning quality, completeness, tone, and domain-specific concerns. 3. **Apply to familiar territory** — Use the checklist on AI output about a topic you know. This lets you calibrate how well the checklist catches real errors. 4. **Apply to unfamiliar territory** — Use the checklist on AI output about a topic you *don't* know well. This exposes gaps in your process — the errors you can only catch with domain knowledge. 5. **Iterate and save** — Update the checklist based on what it missed. Save it in a format you'll actually reach for. **"Done" looks like:** A tested, refined verification checklist (8-12 items) saved in a usable format, with evidence of at least one error it caught and one gap you identified and fixed. --- ## 🧭 Why this matters (Strategists start here) In [EP-Basic-01](/exercises/ethical-prompting/ep-basic-01/), you built a simple 3-question verification prompt. Here, you're building a **systematic process** — a checklist that works regardless of topic, catches both factual and reasoning errors, and is tuned to your specific work context. The community's 75% Ethical Prompting score means most people *intend* to verify AI output but lack a consistent method. A checklist turns good intentions into reliable behavior. At the advanced level, you'll scale this into a governance framework for a team; this exercise builds the individual practice first. --- ## Reflection - Which check caught the most problems? Which was least useful? - How did your verification experience change between the familiar topic and the unfamiliar one? - Is your checklist something you'd actually pull up before sending an AI-generated deliverable? What format makes it most likely you'll use it? - 💬 *Trade checklists with a colleague. Have them apply yours to an AI output from their work — their feedback will reveal blind spots specific to your domain.* (Social Learners) ## ⬆️ Level up Ready for more? Try [EP-Advanced-01](/exercises/ethical-prompting/ep-advanced-01/) — where you'll design an AI governance framework for a team or project. Back to [Ethical Prompting & Judgment](/pillars/ethical-prompting/)