Get to Know AI

The same system can help you see a problem from five angles, then bury the decision under ten pages nobody needs. It can make you faster, expand what you can take on, and improve the quality of your work. It can also move unfinished thinking to someone else's desk.

That range is why people often form an opinion too quickly. One remarkable result and AI becomes magic. One bad result and it becomes a toy. Both reactions mistake one possession for the scouting report.

The better approach is to get to know it.

This guide is for someone who wants to move beyond casual prompting without becoming an AI engineer. It explains what generative AI is doing, where it tends to be strong, where the same strengths can turn into weaknesses, how to give it work, how to keep your judgment active, and how to decide whether the result has earned your trust.

There is no person behind the interface, and the system has no judgment, standing, or accountability of its own. But the habits of good delegation still apply. Give the work a purpose. Supply the context. Set the boundary. Inspect what comes back. Correct it. Decide what happens next.

That is the path from getting an impressive answer to doing better work.

The fast start

If you want to begin before reading the whole guide, choose one real task you already understand. Pick something useful enough to matter and safe enough to get wrong. Give AI this five-line brief:

1. Job: What should come back?
2. Context: What does AI need to know that it cannot guess?
3. Audience and use: Who will use the result, and for what?
4. Boundary: What must it not assume, disclose, promise, change, or decide?
5. Done: What would make the return useful and ready for your review?

Then add three instructions:

• State the important assumptions you made.
• Point me to the evidence behind consequential claims.
• Tell me which part of the answer is most likely to be wrong or incomplete.

When the answer arrives, do not ask only whether it looks good. Ask whether you can explain it, judge it, and trace the parts that matter. Correct one thing and run the task again. The second run often teaches you more than the first.

That is enough to start. The rest of this guide explains why it works.

Part I

What you are working with

AI produces an answer. It does not arrive knowing the answer.

The language models most people use are trained on enormous amounts of text and, increasingly, images and sound. The material is broken into small encoded units called tokens. During training, the model learns patterns that help it predict what token is likely to come next. It is then trained further to follow instructions, behave helpfully, use tools, and solve problems with answers that can be checked.

When you use a model, it reads your prompt and the context available in that interaction. It chooses a likely next token, adds it to the sequence, and does it again. Engineers call that use-time process inference.

This modest-sounding mechanism can produce extraordinary behavior. A model can summarize, compare, translate, draft, classify, explain, write software, reason through a problem, use tools, and revise its own work. Those capabilities are real. So is the underlying fact that the system is generating a response from learned patterns and assembled context.

That explains two things that otherwise seem contradictory.

First, AI can connect clues and fields that no one person could hold at once. Where the learned patterns and supplied context are rich, it can find structure, offer possibilities, and produce useful work at remarkable speed.

Second, AI can fill a gap no one noticed. A missing fact and a missing sentence can look similar to a system built to continue the pattern. The result may sound complete even when the evidence is not.

The practical question is not whether the answer sounds intelligent. Ask: what produced this answer, and what, if anything, certifies it?

The brain in the jar

A useful way to picture AI is as a brain in a clear glass jar. It has three defining qualities.

It is widely read, up to a date

AI can work across more fields than any one person can carry. It can compare documents, explain unfamiliar material, organize evidence, translate specialist language, and give a blank page its first shape. This breadth is one of its greatest strengths.

It is not the same as having every source open in front of it. A model keeps learned patterns, not a perfect library of the pages it encountered. Its training also has a time boundary. Unless the product searches current sources or you supply them, it may not know what changed yesterday, last month, or after its training period.

Use its breadth to map a field. Do not confuse the map with current, local, or accountable expertise.

It is plausible before it is true

AI is very good at producing a response that fits. Often the plausible answer and the true answer overlap. Sometimes they do not, and the language may not warn you.

A fabricated citation can look like a real citation. A weak assumption can appear in the middle of a beautiful analysis. A number can fit the shape of the argument while coming from no authoritative source. The model's tone may remain calm and certain throughout.

It is personally blank

AI does not naturally know your customer, your team's history, the project that failed three years ago, the standard your best reviewer applies, or the concern everyone in the room understands but nobody wrote down. It may remember information you or the product supplied, but that is still assembled context. It is not lived experience.

A person brings relationships, consequences, commitments, purpose, and a history that can surface when a decision requires it. AI brings learned patterns plus the context available now.

A useful shorthand is:

AI answer ≈ learned capability × assembled context

A stronger model changes the first term. A better brief, better sources, better examples, and better tools change the second. Both matter. Neither gives the system a stake in what happens after the answer is used.

Five practical profiles

AI does not have one stable personality, but five profiles help make its working patterns visible. Each profile has a move, a tell, and the human coverage it needs.

These are not model classes. They are ways to watch the work.

The deeper pattern is that AI can become more researched, complete, and convincing without becoming more correct. Its greatest strengths produce the normal signs of quality before the evidence has necessarily earned them.

What stays with people

Good AI use does not depend on pretending people have one mysterious quality called judgment. It helps to separate the human contribution into parts.

Absorption is understanding the return well enough to use and explain it. AI can generate fifty pages. It cannot create fifty pages of attention in your day. If you cannot carry the argument, identify the assumptions, or answer questions about the result, the work has outrun you.

Judgment is deciding what to believe, recommend, reject, prioritize, or own. It draws on evidence, expertise, nuance, relationships, consequences, values, and wisdom. AI can organize those ingredients. It cannot take responsibility for the choice.

Alignment is the human work of turning different understandings, interests, obligations, and authorities into a shared commitment about what happens next. AI can clarify positions and record a decision. It cannot make the commitment on behalf of the people who must live with it.

Part II

Choose the work before you choose the prompt

Ask what you would delegate

The chat window invites the wrong opening question: what can AI do?

That question produces feature lists and demonstrations. A better question wakes up instincts you already have: what would I delegate?

You know that a useful handoff needs context. You know that a new analyst needs a different brief from a trusted veteran. You know that some work can be reviewed against a clear standard and some work depends on taste, local understanding, or authority. You know that handing off the task does not hand off your responsibility for the result.

Those instincts carry surprisingly far.

Start with work that has four qualities:

• You understand the task well enough to notice a bad return.
• A useful result has a recognizable shape.
• The mistake is cheap to find and reverse.
• The task happens often enough for the learning to matter.

Good first tasks include summarizing a document you have read, comparing two versions of a policy, generating options for a presentation you own, reorganizing rough notes, drafting questions for a meeting, finding inconsistencies in a spreadsheet you can check, or challenging an argument you already understand.

Poor first tasks include decisions involving health, rights, employment, legal obligations, material financial consequences, or confidential information you are not authorized to place in the tool. They may eventually benefit from AI, but they require stronger systems, specialist review, and explicit authority.

Delegate, collaborate, or decide

Not every useful interaction is a delegation.

Delegate when the return can be defined and judged. Examples include extracting fields from a document, converting notes into a draft, comparing two lists, applying a rule you have already chosen, or preparing a first analysis against a known standard.

Collaborate when you are still forming the question or the standard. Examples include exploring an unfamiliar problem, developing an argument, considering several strategies, or testing how different people might react. Stay close to the work. Treat the exchange as a thinking process, not an answer service.

Decide when the work reaches a choice no person has made. AI can organize the evidence, state the trade-offs, generate alternatives, and challenge your preference. It cannot decide what your team values, which risk you should accept, whose interests should prevail, or what commitment you are willing to make.

Name the dividend you want

Individual AI tends to produce one of three dividends:

• Better: more evidence, more angles, stronger criticism, or a clearer result.
• More: additional tasks, options, analyses, or versions.
• Time: the same useful work completed with less elapsed effort.

All three are real. They are not interchangeable.

If you want a better decision, another ten pages may not help. If you want time back, producing three new deliverables does not create it. If you want more reach, speeding up a task that was never the constraint may not change the outcome.

Before starting, finish this sentence:

I am using AI here to make this work better / more extensive / faster, and I will know it helped if __________.

Then ask the harder question: what is the outcome actually waiting on?

If your work is the constraint, the gain can travel through. If the next step depends on another person's review, a scheduled meeting, a missing decision, or an unavailable source, AI may only help you reach that constraint sooner.

Know when not to use it

Cheap generation creates an unusual temptation: doing work because it can now be done.

Another analysis, a fourth draft, and a larger set of options can feel productive. Sometimes they are. Sometimes they postpone the decision, consume review time, or create a burden for somebody else.

Skip the handoff when:

• The answer has no clear use.
• You cannot make time to read or judge the return.
• Writing the brief and checking the result will cost more than doing the task.
• The task is primarily the practice you need in order to retain a skill.
• The tool is not approved for the information involved.
• Nobody has the expertise, evidence, or safe test needed to catch the important failure.
• The work would decide something that requires human authority, consent, or standing.

Part III

Brief the work

Prompting is briefing

Prompting often sounds like a technical art built from secret phrases. The useful part is more familiar. You are describing work clearly enough for something else to carry it.

Short prompts are not bad. They are appropriate when the task is small, the stakes are low, and the missing context does not matter. Longer prompts are not automatically better. A five-page instruction can contain three competing goals and bury the one decision that matters.

The five-line handoff

1. Job

Name the return, not merely the topic.

Weak: “Tell me about this market.”

Stronger: “Give me a one-page map of this market that identifies the main customer groups, buying criteria, incumbent approaches, and the three uncertainties that could change whether we enter.”

The second request gives the exploration a job.

2. Context

Supply what the system cannot guess. Useful context may include:

• The decision or situation behind the task.
• Relevant history.
• Definitions used by your team.
• Approved source material.
• Previous attempts and why they failed.
• Examples of good work.
• Facts you already know.
• Assumptions you want held open.

Label the material honestly. A marketing page is advocacy, not independent evidence. A draft policy is not an approved policy. A customer's opinion is not a market fact. AI will often accept the world you assemble for it.

3. Audience and use

Who will receive the result? What will they do with it?

A note for your own exploration can remain provisional. A recommendation for an investment committee needs traceable evidence, stated assumptions, and a much tighter review. A draft customer email needs the customer's situation and the relationship you are protecting.

The use determines the standard.

4. Boundary

State what the system must not do. Examples:

• Do not invent facts, quotes, citations, customer details, or numbers.
• Do not use information outside the attached sources.
• Do not make a legal, medical, employment, or financial decision.
• Do not send, publish, update, delete, purchase, or commit anything without approval.
• Do not treat estimates as actuals.
• Do not name confidential organizations or people.
• Stop and ask when an assumption would materially change the result.

Boundaries become more important when AI can use tools or take actions. An answer can be corrected. A sent message, changed record, or external commitment may have to be undone.

5. Done

Describe what good looks like. You might specify:

• Length and format.
• Required sections or fields.
• Evidence standard.
• Tone and reading level.
• What must be included or excluded.
• The decision the result should make easier.
• The checks that must pass.

A definition of done saves rounds because it turns your private expectation into part of the assignment.

What the five lines look like filled in

Here is the whole brief written out for one real task: comparing two shortlisted payroll vendors before an operating review.

Notice how much of that is history and constraint the system had no way to know. The boundary line is doing the most work: without it, the silence in a proposal comes back as a confident sentence.

A reusable master prompt

Copy this and shorten it for the task in front of you.

Ask for the process you need, not theater

It can be useful to ask AI to show assumptions, calculations, source links, or an outline of its approach. Those artifacts help you inspect the work.

Do not confuse a fluent explanation of its reasoning with direct access to a private mental process or proof that the answer is right. A model can produce a tidy explanation of its own mistake just as it can produce a tidy wrong answer.

Ask for things you can use:

• The facts taken directly from each supplied source.
• The assumptions introduced beyond those facts.
• The calculation, with inputs shown.
• The passages that support each material claim.
• The test cases used.
• The unresolved questions.
• A change log between two versions.

Part IV

Work the relationship

Start wide, then choose depth

AI is often most valuable at the start of unfamiliar work. Let it scout the terrain, surface competing frames, identify likely authorities, and show what is known and unknown.

Then stop and choose what matters.

Decide which uncertainty deserves depth, which source could settle it, whose expertise is required, and what question the work is trying to answer. Return to AI with a narrower assignment.

This sequence is more effective than asking for a comprehensive report and hoping the important point announces itself:

1. Scout the terrain.
2. Identify the questions that could change the decision.
3. Choose the evidence or expertise needed for those questions.
4. Investigate them one at a time.
5. Synthesize toward the decision.

Break ambiguous work into stages

AI handles complex work better when the stages have distinct jobs. A research assignment might become:

1. Define the question and important terms.
2. Map competing explanations.
3. Identify authoritative sources.
4. Extract the evidence.
5. Test the evidence against the explanations.
6. Draft the conclusion with uncertainties visible.
7. Review the conclusion against the original decision.

This is not about turning every task into a bureaucracy. It is about keeping discovery, verification, and judgment from collapsing into one polished return.

For simple work, ask once. For consequential work, create review points before the answer is finished.

Use AI to challenge, not only complete

The easiest way to use AI is to ask it to finish your thought. That is useful and dangerous. The framing inside your question can become the framing inside the answer, and an agreeable system can make your preferred idea feel independently validated.

Build challenge into the interaction:

• “Make the strongest case against my current conclusion.”
• “What would a skeptical customer, operator, regulator, or competitor question?”
• “Which assumption is doing the most work?”
• “What evidence would change the recommendation?”
• “Assume this failed one year from now. What most likely caused the failure?”
• “What am I treating as a fact that is actually an inference?”
• “Offer two genuinely different approaches, including one that rejects my frame.”

Several AI roles agreeing with each other do not create independent evidence. They may share the same model, sources, prompt, and blind spot. Simulated perspectives are useful for generating questions. Real stakeholders are still required where lived experience, standing, or consent matters.

Correct the work and preserve what you learned

Most people use AI as if every interaction were the first. They correct the same tendency repeatedly, restate the same definitions, and rebuild the same context.

A stronger practice treats correction as an asset. When a result misses:

1. Name the failure precisely.
2. Explain why it matters.
3. Provide the corrected source, example, rule, or decision.
4. Ask for a new version that shows what changed.
5. Save the useful instruction, example, or checklist for the next run.

The goal is not one perfect prompt. It is a working method that improves.

What that looks like in one conversation

Most guidance about AI describes the correction. It rarely shows one. Here is a short exchange in full.

Two habits are worth taking from that exchange. Name the failure specifically, because "make it better" gives the system nothing to act on. And say what must not happen, because the invented reason and the invented date were the two most damaging parts of the first answer and neither would have appeared with a boundary line in place.

Project spaces, saved instructions, reusable templates, and documented examples can all carry learning forward. Use the simplest mechanism your approved tool provides. Keep an authoritative source of truth outside the conversation when the information matters.

Learn the tool as well as the task

Working well with AI is a real skill. The entrance is easy and the depth is real, much like a spreadsheet. Almost anyone can type into a cell. Far fewer can build a model other people rely on.

You need two kinds of fluency.

Work fluency is the ability to frame a problem, break it into parts, state assumptions, recognize the evidence needed, judge the result, and connect it to a decision.

Tool fluency is knowing what your AI product can see and do. Can it search current sources? Read attachments? Work across a project? Use a calculator or code tool? Connect to email or files? Preserve memory? Take actions? Show citations? Respect an access boundary?

The first kind determines whether the work is well directed. The second determines which parts the system can carry safely and efficiently.

Expect the frontier to move

AI performance is uneven across tasks and changes over time. A task that fails today may work after a model or product update. A task that worked reliably may change when the model, tool, prompt, sources, or context changes.

Map the frontier against work you understand. Take a completed task for which you hold the answer key and ask AI to do it fresh. Compare:

• Where did it match your work?
• Where did it beat you?
• Where did it confidently miss?
• What did it need from you?
• How much review did it require?
• Did the final result improve the outcome, or only produce more material?

Part V

Verify before you rely

The Verification Boundary

The Verification Boundary is the point at which work has earned its next level of reliance.

Human delegation already works this way. A manager considers who did the work, reviews the underlying analysis, checks important numbers, and watches how someone performs across repeated assignments. Trust grows locally around demonstrated work.

AI deserves the same quality standard with different assumptions. The system may not know what it does not know. It may not flag uncertainty. Its account of why it made a mistake may itself be an inference. Its speed can produce far more material than anyone can responsibly review.

Work earns reliance through at least one of three routes:

1. A credible person can grade it. Someone with the relevant expertise and authority can tell whether the result meets the standard.
2. Evidence or a test can expose the important failure. A source can be opened, a number recalculated, code run, a rule checked, or the cases most likely to break the answer tested.
3. Reality can answer while the stakes remain small and reversible. A pilot, experiment, draft, or contained trial can reveal whether the idea works before it travels farther.

If none of those routes exists, the work may still be useful for exploration. It has not earned the status of an answer.

Check in proportion to reliance

Not every output deserves the same review.

A practical verification routine

Do not outsource verification to the same confidence

Asking the model “Are you sure?” is not a verification method. It may reconsider and catch an error. It may also restate the same answer with greater confidence.

Useful second passes change something material:

• Use an authoritative source the first pass did not have.
• Ask a qualified person who holds a real answer key.
• Run a deterministic calculation or test.
• Start a fresh review with the original sources and explicit failure criteria.
• Ask a different method to reproduce the result.

A second model can help surface issues, but two models may share training patterns, retrieved sources, and common blind spots. Agreement is evidence of consistency, not proof of truth.

Part VI

Manage the weaknesses by name

General warnings such as “be careful with AI” rarely change behavior. Name the failure mode and attach a response.

Each one below leads with what you actually notice, because that is what you will recognize first. The technical name sits underneath, so you have a word for it when you need to explain the problem to someone else.

Part VII

Useful ways to work with AI

The examples below are patterns, not magic prompts. Adapt them to the work and supply the sources needed.

Learn a subject

Use the return to choose what to investigate. Do not treat the first map as expertise.

Read a document deeply

Open the cited pages yourself.

Turn rough thinking into a draft

This keeps structural judgment visible before prose makes everything look settled.

Improve a piece without losing the author

Review the changed meaning, not only the cleaner sentences.

Compare options

Make sure the criteria reflect the real decision, not merely what is easiest to score.

Challenge a recommendation

Then have a person decide whether the challenge changed the recommendation.

Prepare for a meeting

Do not arrive with ten steps when the meeting exists to settle the first one.

Analyze a dataset or spreadsheet

Independently test the calculation and the exceptions.

Design a repeatable workflow

The goal is not maximum automation. It is a useful learning loop.

Part VIII

From private use to shared work

The Delegation Tax

Delegating to AI takes seconds. Doing it responsibly costs attention.

You have to frame the task, assemble context, steer the work, inspect the return, correct mistakes, integrate the useful parts, and explain the result. If the work travels, somebody must understand why it deserves confidence.

The tax does not mean the delegation failed. Human managers pay it too. The question is whether the value of the assignment justifies the attention needed to direct and review it.

Watch three measures:

• How much useful work came back?
• How much human attention did it take to make the result trustworthy?
• Did the work move a real decision or outcome?

Counting outputs while ignoring interventions and downstream review is a poor measure of productivity.

Your private speed does not make the group faster

Individual AI can make one person substantially more capable. The person still has to move the result into shared work.

That creates two crossings. First, the work travels from AI to you. You absorb it and apply judgment. Second, it travels from you to colleagues. They must absorb it in the context of their responsibilities, apply their different expertise and authority, and align on what happens next.

If five people normally expect a week of work to produce one recommendation, that rhythm gives them time to encounter the evidence, ask questions, and form a view. If one person arrives with ten recommendations and twenty next steps, the work can be excellent and still exceed the group's capacity to use it.

Refine before you share

Before sending AI-assisted work to someone else, ask:

The test is not how much material appeared. It is whether the right people can reach a better decision sooner.

Recognize the Personal Ceiling

At some point, the next gain depends less on how well one person uses AI and more on whether the work around that person changes.

You may still be copying context between disconnected tools, remembering why one source was trusted, reconciling versions, and translating a private AI exchange for everyone else. Your attention becomes the integration layer. More personal output creates more claims on that attention.

This is the Personal Ceiling. It is not a reason to stop learning. It is a signal that the next problem is shared context, workflow, standards, review, decision rights, or team design.

Part IX

Build your practice

A seven-day start

A thirty-day progression

At the end of thirty days, do not ask only how often you used AI. Ask what you can now produce, verify, explain, and improve that you could not do before.

Keep a scouting log

For recurring tasks, record:

This turns impressions into a practical map. It also reveals whether the tool is improving the work or merely producing more of it.

Part X

The before-you-send checklist

Before AI-assisted work leaves your hands, check seven things.

The long view

AI is experiential. You will not learn it from a page, including this one.

The people who become effective are not the ones who memorize the largest prompt library. They develop a feel for the work. They know when to let AI run and when to stay close. They recognize the Wide Reader, the Fluent Improviser, the Frame Follower, the Creative Engine, and the Role Player. They know which failures can hide beneath a polished surface. They can brief a task, create a useful review point, and stop when more output is no longer progress.

Most important, they retain intellectual ownership. They use AI to see more, try more, test more, and build more. They do not let production speed decide what matters or persuasive language substitute for evidence. They know that delegation transfers work, not accountability.

The way to begin is not dramatic. Choose one real task. Make the mistake cheap. Give the answer a job. Brief it clearly. Check what matters. Correct the work. Save what you learned. Try again.

That is how a tool becomes a practice.

Sources and further reading

This field guide synthesizes the current Chapters 3 and 4 of Who Holds the Pen?, the longer Chapter 4 working draft, and Mario Suarez's earlier AI teaching materials. The profiles and practice guidance are author-developed teaching models, not scientific classifications.

• National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1, 2024. doi.org/10.6028/NIST.AI.600-1
• Fabrizio Dell'Acqua and colleagues, Harvard Business School Working Paper 24-013 on the jagged technological frontier, 2023. hbs.edu
• Laura Weidinger and colleagues, “Ethical and social risks of harm from Language Models,” 2021. arxiv.org/abs/2112.04359
• Mirac Suzgun and colleagues, “Challenging BIG-Bench Tasks and Whether Chain-of-Thought Can Solve Them,” 2022. arxiv.org/abs/2210.09261
• Laura Jakesch and colleagues, “Co-Writing with Opinionated Language Models Affects Users' Views,” CHI 2023. doi.org/10.1145/3544548.3581196
• Anil R. Doshi and Oliver P. Hauser, “Generative AI enhances individual creativity but reduces the collective diversity of novel content,” Science Advances, 2024. doi.org/10.1126/sciadv.adn5290
• Hao-Ping Lee and colleagues, “The Impact of Generative AI on Critical Thinking,” CHI 2025. microsoft.com
• Judy Hanwen Shen and Alex Tamkin, “How AI assistance impacts the formation of coding skills,” Anthropic, 2026. anthropic.com
• Kate Niederhoffer and colleagues, “AI-Generated ‘Workslop’ Is Destroying Productivity,” Harvard Business Review, 2025. hbr.org
• Jessica A. Reif, Richard P. Larrick, and Jack B. Soll, “Evidence of a social evaluation penalty for using AI,” Proceedings of the National Academy of Sciences, 2025. doi.org/10.1073/pnas.2426766122