No crew pushes back from the gate without a briefing. The captain states the plan, the threats and who does what, so everyone shares the same picture before the engines start. This module teaches your team to brief AI the same way, so the first draft is closer to right and easier to check.
Open the module slides Download the PDF workbook
Time estimates assume reading at about 200 words per minute plus the hands-on exercise. Your pace will vary. Exercises work with any AI chat tool (Grok, ChatGPT, Claude, Gemini or similar). Use non-sensitive data only.
Before every departure, the flying pilot gives a briefing: the runway, the departure route, the initial altitude, the expected weather, what could go wrong and what the crew will do if it does. It usually takes a couple of minutes. It exists because the cost of a shared misunderstanding at 150 knots is far higher than the cost of two minutes at the gate.
Prompting an AI tool is a briefing. A prompt is simply the instruction and information you give the system before it produces anything. And as Module 1 showed, an AI model produces likely text based on what you give it. Give it little, and it fills the gaps with what's typical. Give it a clear picture, and it can produce what's specific to you. Most disappointing AI results trace back to a thin briefing, not a weak tool.
Compare two requests. The first: "Write an email to a customer about a delay." The second: "You're writing for a family-owned roofing company. Draft a short, apologetic email to a homeowner whose job is delayed three days because of rain. Offer two new start dates (Tuesday or Wednesday), keep it under 120 words, warm but professional, no discounts promised."
The first will produce a generic email that someone must rewrite heavily, and that might promise something your company doesn't offer. The second will usually produce something close to sendable. Thirty extra seconds of briefing saves several minutes of editing and lowers the chance of an error slipping through.
There's a second, less obvious benefit. A clear briefing sets the standard you verify against. If you told the AI "under 120 words, two dates, no discounts," you know exactly what to check. A vague prompt gives you no yardstick, so you're left judging whether it "seems fine," and Module 1 showed where that leads.
In a cockpit, briefing is a habit, not an occasional extra. It happens every flight, even on the hundredth departure from the same airport, because familiarity breeds assumptions. Teams that get consistent value from AI treat briefing the same way. They don't save good prompting for big tasks. It becomes the normal way to ask.
Ethan Mollick's advice to work with AI as you would with a capable colleague fits here. You wouldn't hand a new team member a one-line note and expect a perfect proposal. You'd explain the customer, the goal, the constraints and what good looks like. The AI is the same, except it has no memory of your business unless you provide it, and it will never ask for clarification unless you invite it to.
That last point leads to one of the simplest, most effective techniques: let the AI ask you questions first. Add a line such as "Before you start, ask me up to three questions about anything you need to know." It's the equivalent of the first officer asking "Confirm we're expecting the visual approach?" before anyone commits. It surfaces missing information early, while it's cheap to fix.
Lesson 2 gives you a five-part briefing structure. Lesson 3 covers examples and source material. Lesson 4 covers iteration, the conversation that turns a draft into a finished product. Lesson 5 shows how to save good prompts as team checklists, so the whole crew benefits.
Example: a veterinary clinic's reminder texts. A receptionist asks an AI tool to "write appointment reminder texts." The results are friendly but include a link to "manage your booking online," which the clinic doesn't have, and run too long for a text message. With a briefing that names the clinic, the 160-character limit, the call-to-confirm process and the tone, she gets usable reminders on the first try and knows exactly what to check.
Exercise (about 7 minutes). Pick a real but non-sensitive message you need to write this week. First, ask your AI tool in one vague sentence and save the result. Then write a full briefing: who you are, who it's for, the goal, the limits and the tone, and end with "Ask me up to three questions before you start." Answer its questions, then compare the two outputs. Note how many edits each would need before sending.
Aviation briefings follow a standard order, so nothing gets missed and everyone knows what's coming. A consistent structure for AI prompts does the same. The one we teach has five parts. You won't need all five for every small request, but running through them catches most of what causes poor results.
Tell the AI what perspective to write from: "You're an experienced customer-service lead at a regional HVAC company" or "Act as a careful editor who specializes in plain-language writing." Mollick points out that a role is a quick way to give context, because it shifts the style, vocabulary and assumptions of the answer. A role isn't magic. It doesn't give the AI expertise it lacks. But it steers the output toward the right register.
State the job with a clear verb: draft, summarize, compare, list, rewrite, critique, extract. "Help with this proposal" is vague. "Rewrite the opening paragraph of this proposal so it leads with the customer's problem" is a task. If there are several steps, number them.
This is the part people skip, and the most important. Who's the audience? What's the situation? What has already happened? What matters most? What's true about your business that a generic answer would get wrong? Context is where you replace the AI's assumptions with your facts.
Length, reading level, things to include, things to avoid, and rules: "Don't promise discounts." "Use only the facts below." "If you're unsure, say so instead of guessing." "No medical advice." Constraints are your guardrails. They're also your verification checklist afterwards.
Email, bullet list, table, script, three options to choose from, a draft followed by a list of assumptions. Asking for the format you'll actually use saves reformatting. Asking for "a list of any assumptions you made" at the end is especially useful, because it shows you where the AI filled gaps.
Here's a full briefing for a common small-business task:
Role: You're an office manager at a five-person bookkeeping firm. Task: Draft an email reminding clients to send year-end documents. Context: Our clients are small local businesses. The deadline is January 31. Documents can be uploaded to our secure portal or dropped off at the office. Last year many clients sent documents late, which delayed their returns. Constraints: Under 150 words. Friendly, not scolding. Don't mention fees or penalties. Don't invent portal features. Format: Subject line plus email body, then a short list of anything you assumed.
That took about a minute to write. The result will usually need only light edits, and the constraints tell the reviewer exactly what to check.
A captain's briefing for a routine departure in good weather is shorter than one for a challenging airport at night in a storm. Scale yours the same way. A quick rewrite might need only task and constraints. A customer-facing document, a policy or anything with financial or legal weight deserves all five parts, plus source material, which is the subject of the next lesson.
Pilots use memory aids for critical sequences. For prompts, try "R-T-C-C-F" or simply ask yourself "Who, what, why, limits, look?" before you hit enter. Within a week or two it becomes automatic.
Example: a plumbing company's job posting. The owner's first attempt, "write a job ad for a plumber," produces a generic posting listing benefits the company doesn't offer. With the five-part briefing (role: hiring manager at a 12-person residential plumbing company; context: real pay range, schedule, licence requirement, what makes the team good to work for; constraints: no invented benefits, under 250 words; format: headline, summary, bullets), the draft is accurate and on-brand, and checking it takes two minutes against the facts he supplied.
Exercise (about 9 minutes). Choose a task you do at least monthly. Write a five-part briefing for it using the R-T-C-C-F structure, and include the line asking for "a list of any assumptions you made." Run it in your AI tool. Then check the output against each of your constraints, one by one, and read the assumptions list. Did the AI assume anything wrong? Add that missing fact to your context and run it again.
Pilots don't fly an approach from memory. They brief from the published chart for that runway at that airport, because the details that matter (altitudes, frequencies, minimums) are specific and unforgiving. The chart is the authoritative source, and everyone flies from the same one.
When accuracy matters, AI needs a chart too. The single most effective way to get accurate, specific output is to give the AI the actual source material and tell it to work from that. This is often called grounding.
Instead of asking the AI what your return policy is, paste in your return policy and ask it to draft a customer reply based on it. Instead of asking for a summary of "standard" contract terms, provide the contract. Instead of letting it guess your prices, give it the price list. Then add an instruction such as:
"Use only the information provided below. If the answer isn't in it, say 'Not covered in the material provided' rather than guessing."
That one line changes the AI's job from "produce a plausible answer" to "work from this document," and gives it explicit permission not to fill gaps. It doesn't guarantee accuracy, so you still verify, but it reduces invented details considerably, and it makes verification faster because you know exactly where each fact should come from.
Mind your data rules from Module 1. Grounding means sharing material with the tool, so only paste documents your company allows in that tool, and strip personal or confidential details when you can.
A useful addition is to ask the AI to point to where each key fact came from: "After each claim, quote the sentence from the source that supports it." That's like a crew member calling out the chart reference as they set an altitude. Unsupported claims stand out immediately, and the reviewer can check quotes quickly.
The second powerful technique is giving examples of what good looks like. If you want customer replies in your company's voice, include two or three replies you've written before that you're proud of. If you want product descriptions in a certain format, show one. AI tools are very good at matching a pattern they're shown, often better than following an abstract description of a style.
A few tips for examples:
Anthropic's engineering guidance notes that for many business applications, a single well-crafted request with relevant retrieved information and good examples is enough. You often don't need anything more elaborate. That's a useful reminder for small businesses: before reaching for complex tools, get grounding and examples right.
Sometimes you're asking about something you don't have a document for, such as a regulation or a general business question. In that case, treat the output as a starting point for research, not an answer. Ask the AI to name the official source you should check (a government agency, a standards body, the vendor's documentation), then go and read it. For anything with legal, tax, safety or medical implications, the official source and a qualified professional have the final word.
Example: a fitness studio's membership FAQ. The studio's first AI-written FAQ invents a "freeze your membership for up to three months" policy. The owner tries again, pasting in the actual membership agreement with the instruction "use only this document; say if something isn't covered," plus two existing FAQ answers as style examples. The new FAQ is accurate, matches the studio's voice, and flags two common questions the agreement doesn't answer, which turns out to be useful for updating the agreement.
Exercise (about 8 minutes). Take a short, non-confidential document from your work (a policy, a service description, a price sheet, or a made-up one). Ask your AI tool a customer question about it without pasting the document, and save the answer. Now paste the document with the "use only the information provided" line and the instruction to quote supporting sentences, and ask again. Compare: what did the first answer invent? Was anything in the second answer unsupported by a quote?
No flight goes exactly to plan. Winds shift, air traffic control reroutes you, the weather at the destination changes. Professional crews don't treat a change as a failure. They make small, deliberate corrections and confirm each one. Getting from takeoff to landing is a series of adjustments.
Working with AI is the same. The first response is rarely the final product, and it isn't supposed to be. The real skill is iterating: giving clear feedback, adjusting the brief and steering the output to where you need it. Mollick has noted that people often use AI in a shallow, one-question-one-answer way and miss much of its value, which comes from back-and-forth.
When a draft misses, resist typing "that's not right, try again." The AI doesn't know what was wrong. Instead, give feedback the way a good instructor debriefs a student: specific, actionable, one or two points at a time.
A handful of moves cover most situations:
That critique-then-revise loop is a simple version of what Anthropic's engineering guidance calls an evaluator-optimizer pattern, where one step produces a draft and another evaluates it against criteria, repeating until it's good enough. You can run it by hand in any chat tool.
Iteration has a limit. If you've made four or five corrections and the output still isn't right, stop and ask why. Maybe the AI lacks information only you have. Maybe the task sits outside the jagged frontier from Module 1. Maybe it's simply faster to write the last 10% yourself. Pilots disconnect the autopilot when it's quicker and safer to hand-fly. Do the same. The goal is a good result, not proving the AI can do it alone.
In long conversations, an AI can drift: it forgets an early constraint, reintroduces something you removed, or contradicts an earlier version. Before you use the final output, re-run your original constraints as a checklist against it. Every iteration is a chance for a new error, so the final check still applies, however good the conversation felt.
Example: a catering company's proposal. A catering manager's first AI draft of an event proposal is accurate but bland. Instead of starting over, she asks for three alternative openings, picks the warmest, asks the AI to critique the draft "as a busy corporate event planner," and has it tighten the menu section based on that critique. Four focused turns produce a proposal she's happy with. Her final check catches that the AI had quietly changed the guest count from 85 to 80 in one turn, so she corrects it before sending.
Exercise (about 8 minutes). Take the output from Lesson 2's exercise (or a new draft). Run at least four deliberate iteration moves: ask for three options for one section, ask the AI to critique its draft from your customer's point of view, revise based on that critique, and lock one part while changing another. Finally, check the end result against your original constraints. Did anything drift during the conversation?
Every checklist in a cockpit started as hard-won knowledge: somebody learned, sometimes painfully, that a step mattered. Writing it down turned one person's lesson into everyone's standard. That's why a new first officer can fly safely with a captain they've never met. They share the same procedures.
Good prompts deserve the same treatment. When someone on your team works out a briefing that reliably produces good customer replies, estimates or summaries, that knowledge shouldn't live in one person's chat history. It should become a shared, standard briefing the whole crew can use and improve. We call that a prompt library.
Keep each entry short and practical, like a checklist card:
The template is the heart of the entry. A good one makes the right thing easy:
You're a customer-service lead at [company]. Draft a short email to [customer first name] about a change to their [service] appointment. Original date: [date]. New options: [option 1] or [option 2]. Reason: [reason, in one plain sentence]. Constraints: under 120 words, apologetic but upbeat, no discounts or credits unless I list one here: [none]. Format: subject line plus body, then list anything you assumed.
Anyone on the team can fill in the blanks in under a minute, and the reviewer knows exactly what to check.
Simple beats sophisticated. A shared document or spreadsheet that everyone can find is enough for most small and mid-size teams. Some AI tools let you save custom instructions, reusable prompts or team assistants. Those can be useful, but keep a plain master copy too, so your procedures don't depend on one product. Remember the training half-life from Module 1: tools change, and your library should survive a tool change.
Checklists get revised when procedures change or when someone finds a better way. Treat your library the same:
A template makes good results more consistent. It doesn't make them guaranteed. Familiarity with a checklist can create its own complacency, the "I've run this a hundred times" effect. The verification steps in each entry are not optional, especially for anything a customer, regulator or the books will see.
You now have the full preflight: brief clearly (Lessons 1–2), fly from the chart with grounding and examples (Lesson 3), make deliberate course corrections (Lesson 4), and turn what works into shared standards (this lesson). Module 3 moves from how to brief to what to hand AI in the first place.
Example: an insurance agency's shared templates. A six-person independent insurance agency found each agent prompting AI differently, with uneven results. They collected the best prompts into a shared document with eight templates (renewal reminders, coverage-explanation drafts, meeting summaries and so on), each with verification steps and a clear note never to paste policy numbers or personal details into the consumer tool. New staff now get up to speed faster, and the office manager reviews the library every quarter.
Exercise (about 8 minutes). Turn your best prompt from this module into a library entry. Write the name and purpose, the five-part template with [brackets] for the parts that change, what to paste and what never to paste, three or four verification steps, a traffic-light rating, and yourself as owner. Then test it: fill in the blanks for a new situation and run it. Did the template hold up? Adjust it before saving.
All lesson text is original and summarizes these ideas in our own words. We link to the originals and encourage you to read them.
Business examples are illustrative composites, not real clients. Nothing here promises any specific productivity, savings or financial result. Not legal advice.