Exercises — Chapter 29: Writing with AI

Writing is learned by writing — and writing with AI is learned by doing the thinking yourself and judging the tool's output. Most of these require you to produce or revise real text, or to catch a tool's error. Where a task is open-ended, a self-assessment rubric follows in place of an answer.

Setup note. You can do every exercise with or without access to a live AI tool. Where an exercise says "ask the model," you may instead use the AI outputs provided in the prompt (they are illustrative composites, realistic of current general-purpose models) and reason about them. The skill being practiced is judgment about AI output, not access to any particular product.

Part A — Analyze This ⭐

Read each AI output or scenario and identify what works and what's broken. The skill: telling fluent-and-useful from fluent-and-empty (or fluent-and-wrong).

A1. Here is an AI-generated opening for a blog post about a database migration. Name three specific things wrong with it, using vocabulary from this chapter and Chapter 3.

"In today's fast-paced, data-driven world, database migrations have become more important than ever. This revolutionary process empowers organizations to seamlessly unlock the full potential of their data infrastructure. Let's dive in and explore the exciting journey of modernizing your data stack!"

A2. A teammate pastes this AI fact into a report draft. Which words tell you it needs verification before it can stay, and why?

"According to a 2020 MIT study, teams using asynchronous communication completed projects 31% faster, and 92% of surveyed developers reported higher job satisfaction."

A3. Two prompts produced two outputs for the same task ("explain our API rate limit to developers"). One is generic, one is specific. Identify which is which and name the prompt feature that most likely caused the difference. - Output 1: "Rate limiting is an important technique for managing API traffic. It helps ensure fair usage and system stability for all users." - Output 2: "You get 100 requests per minute per API key. Hit the limit and you'll get a 429 with a Retry-After header telling you how many seconds to wait. Batch endpoints count as one request, not one per item."

A4. Classify each use of AI below as augmentation (revising your own thinking) or outsourcing (the model supplying the substance): - (a) "Here's my conclusion paragraph — make it tighter." - (b) "Write the conclusion for my essay on supply-chain resilience." - (c) "Argue against my thesis so I can find the weak points." - (d) "Summarize this 40-page report so I don't have to read it" (and you never read it). - (e) "Give me ten title options for this article I wrote."

A5. This AI output for a performance-review summary sounds complete. Why is it structurally impossible for the model to have written this well, and what's the integrity problem if a manager submits it?

"Jordan consistently exceeded expectations this quarter, demonstrating exceptional leadership on the payments project and strong collaboration across teams. Their attention to detail and proactive communication made them an invaluable team member."

A6. A student's prompt history on a submitted essay: (1) brainstorm topics, (2) outline the argument, (3) "improve the flow of my three body paragraphs," (4) "write a stronger, more persuasive conclusion." The course policy permits AI "for idea generation and feedback only." Identify the single step that most clearly violates the policy and explain why.

A7. Here are two summaries of the same study by an AI. The study "found a modest, statistically significant association between X and Y in one age group." Which summary is dangerous, and what specific Chapter 7 failure does it commit? - Summary 1: "The study found a modest association between X and Y in one age group." - Summary 2: "The study proves that X causes Y."

A8. A developer says: "I used AI to write this function in Rust. I've never used Rust, but the code looks clean and it compiled, so I shipped it." Apply the governing rule (§29.7). Was this a safe use? What's the gap?

Part B — Revise This ⭐⭐

Improve weak material. The scenario is given, not the answer. These center on the chapter's two signature tasks: improving a weak prompt, and fixing a fluent-but-flawed AI draft.

B1. Improve this weak prompt. A project lead types this and gets bland mush back:

"Write an email about the deadline change to the team."

Rewrite it as a real brief using all four levers (§29.5): role, audience, constraints, and (invent a plausible one) an example of the team's voice. Supply realistic specifics: the deadline moved from June 14 to June 28, the reason is a delayed dependency from another team, and the lead wants to be honest without throwing the other team under the bus.

B2. Improve this weak prompt (academic). A student types:

"Help me write my literature review on coral bleaching."

Rewrite it so the model assists the student's own thinking rather than generating the review (which would violate most policies and defeat the assignment). Hint: what could the student legitimately ask the model to do — given a draft the student wrote — that stays inside "feedback and revision"?

B3. Fact-check and fix this fluent-but-wrong AI draft. Below is an AI-generated paragraph for a report. At least three claims are the kind that must be verified or cut. Mark each suspect claim, say why it's suspect, and rewrite the paragraph keeping only what a careful writer would publish without a source (and noting what you'd verify).

"The first email was sent by Ray Tomlinson in 1971, and within just two years, email accounted for 75% of all ARPANET traffic. A landmark 2018 study from Stanford's Dr. Helena Marsh found that the average knowledge worker now spends exactly 28% of their workweek on email, and that reducing email volume by half increased productivity by 41%. Email remains a cornerstone of modern communication."

B4. De-genericize. Here is an AI first draft of a README intro for an open-source date-parsing library (Raj's chronoparse, the running example from Chapter 24). It's fluent and empty. Rewrite it to be specific and honest, inventing realistic concrete details (what it does, the one thing that makes it useful).

"chronoparse is a powerful and flexible library that makes working with dates easy and intuitive. With its robust feature set and seamless integration, chronoparse empowers developers to handle dates effortlessly. Get started today and unlock the full potential of date parsing!"

B5. Restore the voice. A writer fed their blunt, plain-spoken Slack announcement to an AI to "make it more professional," and got back the version below. The writer's actual voice is direct and human. Rewrite the AI version back toward a real, direct voice — keep it professional but kill the inflation.

"We are pleased to inform the team that we will be undertaking a planned maintenance window during which certain services may experience temporary unavailability. We appreciate your patience and understanding as we work diligently to enhance our infrastructure."

B6. Fix the over-hedged AI draft. Asked to state a clear, well-supported finding plainly, the model produced mush. Rewrite it to state the finding with appropriate (not excessive) confidence. The underlying fact: the new caching layer cut median page-load time from 1.2s to 0.4s in production over two weeks.

"It could potentially be suggested that the caching layer may have possibly contributed to what might be interpreted as some degree of improvement in page-load times, although results may vary and further investigation could be warranted."

Part C — Write This ⭐⭐–⭐⭐⭐

Scenario → produce the document. Do the thinking yourself; use AI only as the chapter prescribes (and note where you would).

C1. You're announcing a small feature to your users (invent the feature and audience). First, write a rough draft in your own words, with your own substance. Then write the prompt you'd use to have a model help you revise it (role, audience, constraints, example). Submit both the draft and the prompt. (You do not need to run the prompt — the point is to demonstrate the correct order of operations: think and draft first, brief second.)

C2. Write a short AI-use disclosure statement for a research paper, in two or three sentences, of the kind a journal might require — honestly describing a realistic, legitimate use (e.g., "the authors used a large language model to improve the clarity and concision of selected passages; all factual content, analysis, and conclusions are the authors' own and were verified by the authors"). Then write a one-sentence note on why the second clause (the verification/ownership claim) matters.

C3. A junior colleague asks you: "When is it OK to use AI at work?" Write them a clear, kind, four-to-six-sentence reply that gives them the governing rule (§29.7) and the augmentation/outsourcing distinction (§29.4) in plain language — no jargon, the way Chapter 28 would explain it to a general audience.

C4. Take a paragraph you've actually written recently (any source). Write the prompt that would ask a model to steel-man the opposite of your point so you can pressure-test it (§29.4). Then write two sentences on how you'd use the model's counter-argument — specifically, how you'd keep the judgment about whether it's right with yourself.

C5. Write a 120–180 word section for a team's internal wiki titled "Our policy on using AI for writing." Invent a reasonable policy that reflects this chapter's principles: what's encouraged (revision, brainstorming), what requires care (verification of facts), what's not allowed (passing off AI-generated substance, anything you can't evaluate), and the verification expectation. Make it specific and usable, not generic.

Part D — Synthesis & Critical Thinking ⭐⭐⭐

Cross-chapter integration, finding the flaw, translating for audiences.

D1. Find the flaw. A productivity guru posts: "Stop wasting time writing first drafts. Let AI write the draft, then you just edit it. You'll write ten times faster." Using Chapter 1 and §29.4, explain precisely what's wrong with this advice — not "AI is bad," but the specific thing this workflow skips and why it matters.

D2. Translate for three audiences. Take this single technical claim and write it three ways, then say which (if any) you'd trust an AI to help draft and why: "Our LLM-based feature has a non-trivial hallucination rate, so all outputs are reviewed by a human before reaching the user." Rewrite for (a) a fellow engineer, (b) a non-technical executive deciding whether to fund it, (c) an end user reading a help page. (Ties Chapter 2, Chapter 28.)

D3. The integrity case. A data scientist uses AI to help write the prose of a memo (Dana Whitfield's churn memo, the running example) but does all the analysis herself and verifies every number. A second data scientist uses AI to do the analysis (interpret the data and draw the conclusions) and writes the prose herself. Both memos read identically. Which use is fine and which is the problem, and why does "the memos read identically" not settle it? (Ties Chapter 11, §29.6.)

D4. The governing rule, generalized. The chapter's rule is "if you can't evaluate the output, don't use AI for that task." Argue whether this same rule applies to any tool you don't understand (a statistics package you can't interpret, a translation into a language you don't read, a chart from a library you can't sanity-check). Is the rule really about AI, or about something more general? Defend your answer in a short paragraph.

D5. The thinking trade-off. Chapter 1 said writing surfaces gaps in your understanding because the page forces the logical joints you'd skip in your head. If a model writes the connective prose for you ("therefore," "because," "this implies"), what specifically do you risk not discovering? Connect this to why the chapter insists you draft before the model touches the work.

Part M — Mixed Practice (Interleaved) ⭐⭐–⭐⭐⭐

These mix this chapter with earlier ones, so you must choose the right approach — and decide whether AI even belongs in the task.

M1. You're writing a docstring (Chapter 24) for a function you wrote, and you ask an AI to draft it. The AI produces a docstring that describes the parameters correctly but states the function "returns the user's age" when it actually returns a date of birth. Name two chapter principles that catch this error (one from Chapter 24, one from this chapter) and say what you do.

M2. A peer-review situation (Chapter 12): a colleague asks you to review their report, and you're tempted to paste it into an AI and forward the AI's critique as your feedback. Using Chapter 12's "give feedback that helps" and this chapter's §29.4/§29.6, decide whether this is acceptable, and under what conditions (if any) it could be.

M3. You need to cite a source (Chapter 11) and an AI gives you a perfectly-formatted IEEE reference for a paper that supports your point. What does Chapter 11 and this chapter require before that citation can go in your document, and why is the formatting being perfect irrelevant (even a warning sign)?

M4. You're translating a churn finding for a VP (Dana Whitfield's task, Chapter 2 / Chapter 27). You ask an AI to "make this executive-friendly." It produces something fluent. Which two things must you still supply that the model cannot, and how do you tell whether its "executive-friendly" version actually fits your VP?

M5. An email delivering bad news (Chapter 19): you draft it, then ask AI to "soften the tone." The result is so softened the actual message is buried. Name the Chapter 19 principle violated and the Chapter 7 principle, and decide how you'd use the AI's output (if at all).

M6. You're writing instructions (Chapter 22) and ask AI to generate troubleshooting steps for a device. You've never used the device. Apply the governing rule (§29.7): can you use the AI's steps? What's the specific danger if you publish steps you can't verify, and how does it connect to Chapter 22's "someone who has never done this" test?

Part E — Extension ⭐⭐⭐⭐

For motivated readers and the Deep Dive track.

E1. The hallucination hunt. Find (or recall) a real instance where an AI gave you a confident, specific, wrong answer — a fake citation, a wrong fact, an invented detail. Document it: what it said, how you caught it (or would have), and what would have happened if you'd trusted it. Then write a one-paragraph reflection connecting it to §29.1 (why it happened) and §29.7 (why the rule would have protected you). If you've genuinely never caught one, that itself is worth examining — does it mean the tool is reliable, or that you haven't been verifying?

E2. The voice experiment. Take a paragraph in your own genuine voice. Feed it to an AI three ways: (a) "make it more professional," (b) "make it more engaging," (c) "rewrite it." Compare the three outputs to your original. Write a short analysis: what did each instruction do to your voice, and which (if any) preserved it? What does this tell you about §29.5's claim that examples beat descriptions for controlling voice?

E3. Draft a real AI policy. Write a complete, usable AI-use policy (250–400 words) for a context you know — a course, a team, a publication. Cover: permitted uses, required disclosure, verification expectations, forbidden uses, and the reasoning behind each (don't just rule — explain, the way this book does). Pressure-test it against the edge cases in Part A6 and Part M2: would your policy give a clear answer in those gray zones? Revise until it does.