AI on a real project: a working case study

Three people are about to read about you, and no two will read the same way. A recruiter opens your resume with dozens more waiting and may give it less than a minute. A hiring manager reads your cover letter slowly, hunting for evidence that you understand the actual work. A former coworker glances at your two-line note on a networking site and decides on the spot whether to answer. One person. One true story. Three readers, three formats, three completely different goals.

Most job seekers write one document and send it everywhere, which is why most applications read as if written for no one. You already own every tool it takes to do better. Lesson 2 taught you to ask well. Lesson 3 taught you to build an assistant that knows your background before you type a word. Lesson 4 taught you to judge which tasks deserve a handoff at all. Each skill arrived separately, with its own practice. This lesson runs them all at once, on one project with real stakes, from first draft to final handshake. Nothing new arrives today. The tools stop being a collection and start being a method.

This lesson adapts the session that closes the using-AI-well unit of the Harvard Kennedy School course this track is built on, created by Sharad Goel, Dan Levy, and Teddy Svoronos. That session is a case study: everything the unit taught, prompt anatomy, tailoring, the decision criteria, run end to end on a single communications campaign. It ends with an invitation: take the method to a situation “closer to what you do in your day-to-day life.” This lesson accepts the invitation. Our project is a job search, with a detour to a small nonprofit near the end, and the method survives the move without a scratch.

One story, many readers

Start where the session’s own activity starts: with the shape of the task. Before anything goes to the machine, it has you pick the platform you are writing for, aim each message at a different group of readers, and vary what each message is for. Hold those three choices up as questions and you can ask them before any piece you ever produce. Who will read this? Where will they read it? What should happen because they read it? Audience, platform, goal. It sounds almost too simple to matter. It is the whole method.

Run it on the job search. The resume bullet’s audience is a recruiter; its platform is a skim, quite possibly filtered through screening software first; its goal is survival to the next round. So it stays short, factual, dense with the posting’s own vocabulary. The cover letter faces the hiring manager, gets an actual read, and must earn an interview. So it can breathe, connecting two or three real wins to problems this team visibly has. The outreach note goes to someone who already knows you, lands between meetings, and wants a conversation, not a job. So it stays warm and asks for fifteen minutes, nothing more. The facts never change. The shape changes every time, on purpose.

Before you hand any of this over, notice how loudly this task passes lesson 4’s first filter. The task needs personalization: the same story tailored to different readers. There is a body of text to draw from: your old resumes, your notes on past wins, the job posting itself. It rewards creative variation. And you hold demonstration data: the strongest bullet you ever wrote is a worked example of your standard. The course runs this same scoring on its own project and reaches the same verdict. This is the machine’s kind of work.

Now the handoff. The assistant you built in lesson 3’s practice already holds your resume and standing instructions, so you skip the introductions and go straight to work: task, instructions, context, exactly as lesson 2 drilled. When you are not sure what context the machine needs, use the session’s best tip: turn the question around. Its teacher, unsure what to feed the tool, stated the goal and added, “please ask me what you need to know to do this task well.” Back came a list of clarifying questions, several the teacher admitted to never having considered, and answering them made the drafts markedly better. Ask the machine to interview you about the task before it does the task. It costs one sentence.

Then iterate, because a first draft is a proposal, not a verdict. Read each draft the way its reader will: the bullet in seconds, the letter with a skeptic’s eye. Say what is off and let it revise. Draft, react, sharpen, again. Speed is what you handed over. Judgment stayed home.

Which pieces stay yours

Lesson 4’s second filter, consequences, decides which pieces the machine may own. Drafting: yes, all of it, variations and rewrites included. But every claim about you that survives into a final document gets your check, line by line, against what actually happened. That rule has teeth. The machine is a prediction machine, and confident application language is exactly what its training text is full of. Left unwatched, it will quietly promote you. Helped with becomes led. Familiar with becomes expert in. A two-person effort becomes yours alone. It is not lying. It is predicting. You are the only fact-checker your own life has, so never let the machine invent an accomplishment.

Privacy rides along too. Your story is yours to share. Your current employer’s is not. A draft will sometimes reach for confidential material to make a bullet land: revenue figures, client names, anything you hold in trust. Swap in figures you genuinely have the right to share, or shapes instead of numbers: grew, halved, doubled. Lesson 4’s question, asked before pasting, settles every case: am I comfortable with this information leaving my hands under this tool’s rules?

Rehearse before the room

Materials get you the interview. The session’s most surprising move is what comes next: use the machine to rehearse for it. Many of the course’s own students had never considered preparation and simulation until the case put it in front of them. The setup takes one prompt. Tell your assistant to act as the interviewer for this role, and be explicit about the format: do not write the whole exchange as a script; ask one question, then stop and wait for my answer.

This is lesson 2’s persona move carrying real weight. Because the assistant holds your resume and the posting, its questions are not generic. They aim at your actual seams: the two-year gap, the tool the posting names that your resume does not, the career change you have never practiced explaining. And the role-play takes direction. The session notes you can tell the tool to play its part more aggressively or less, tuning the exchange into exactly the kind of preparation you want. For you that means a friendly screening call on the first night and a skeptical, pressing panel the night before the real thing. When an answer comes out weak, say so and run it again. Ask what a doubtful interviewer would push on next. Rehearse the follow-up, not just the opener.

Be honest about what this rehearsal is, because the course is. Its own classroom split. Some students found the simulation genuinely capable. Others said a live spoken performance is mostly voice and presence, and a text exchange sits a long way from that. Both are right, and the session’s teacher sides with caution: this is far from the real event, and nobody is proposing to remove the humans from your preparation. What the machine coach offers is different. It is available at midnight, needs no scheduling, and never tires of your seventh attempt. Use it as a complement: let it find the questions you have no answer for while discovering them costs nothing, and keep one live mock interview with a human who can hear you.

Change the mission, keep the method

Now watch the method travel, because nothing in it knew it was about a job search. Picture a three-person food pantry that needs donors and volunteers. One true story: what the pantry does, whom it serves, what a donation becomes. Three readers: a foundation officer scoring a grant application against a rubric, neighbors scrolling a community page, a store owner deciding whether to sponsor a shelf. Audience, platform, goal, named before every piece. The body of text to draw from is the pantry’s own documents, with the same privacy check: volunteers’ and clients’ stories go in only with their consent. And the night before the funder meeting, the director rehearses against a skeptical program-officer persona that has read the application and pushes on the weakest number. A job seeker and a tiny nonprofit just ran the identical method. So will whatever project you are actually holding.

Where help ends and deception begins

One conversation remains, and the course is honest enough to have it out loud. The session sorts uses of AI along a spectrum. Some sit comfortably: preparing, rehearsing, drafting from your own material. Some leave some people uneasy. And some, the session says plainly, most people would find hard to defend. A job search has the same spectrum, and you should know where the line sits before you are tired and tempted.

Here is the line, and it is brighter than it first looks: after reading, does this person believe anything false about you? Polishing true sentences leaves every fact standing; that is the service good editors and career counselors have always sold. Fabrication plants a false belief: experience you do not have, a skill level you cannot demonstrate, a take-home assignment submitted as your own unaided work when the employer asked to see yours. The practical case against it is as strong as the moral one. Application materials are promises that the interview collects on. Fabricate a line and you have scheduled your own exposure, in the one room the machine cannot enter with you.

Two habits keep you on the right side. First, patrol the drafts for inflation you did not ask for. Those quiet promotions are how honest people drift across the line one plausible word at a time; strike anything you could not defend under a follow-up question. Second, default to daylight. When the course’s classroom debated remedies for AI-made deception, the first answers on the board were transparency and disclosure. Your version is simpler: if an employer asks whether or how you used AI, answer plainly. And if the thought of a reader knowing how a piece was made makes you flinch, the flinch is your answer about the piece.

This lesson’s practice runs the whole method in Clawless, on your project: name a real target role (or borrow the pantry), produce two pieces shaped for two different readers, make one explicit which-pieces call with lesson 4’s filters, then rehearse with the interviewer persona until it finds a question you cannot answer yet. The assistant you built in lesson 3’s practice is the right starting point. It already knows your background.

Why this matters when you use AI

The method is not about job hunting. One project, many readers, judgment at every handoff, rehearsal before the stakes: that skeleton fits a grant, a pitch, a proposal, a hard announcement. You now recognize the shape of AI-assisted work done well.
Rehearsal may be the most underused skill in this track. Most people only ask the machine to produce things. Asking it to push back, in persona, at midnight, before a moment that counts, is where the confidence comes from.
One question keeps every use defensible. Would this still work if the reader knew how it was made? Polish survives that question comfortably. Fabrication does not survive it at all.

Common pitfalls

One document for every reader. If two different audiences received the same text, you skipped the three questions, and the method never started.

Accepting the flattering draft. The machine inflates by default, because impressive language is probable language. The bullets that please you most are precisely the ones to check first.

Rehearsing only in text. The machine coach finds your weak answers; it cannot hear your voice shake. Keep one live mock conversation with a human in the plan.

Feeding it what is not yours to share. Your employer’s confidential numbers and other people’s stories do not become shareable just because the draft lands better with them in it. Ask lesson 4’s privacy question before every paste.

What you should remember

The method fits in one sentence: name the audience, the platform, and the goal for every piece, then vary the shape while the facts stay fixed.
Your lesson 3 assistant is the right vehicle: standing instructions plus your documents, written once, so every draft starts specific.
When you do not know what context to give, ask the machine to ask you. One sentence buys a list of questions you would not have thought of.
Draft with the machine, verify as yourself: every final claim about you gets a line-by-line check, and quiet inflation gets struck.
Rehearsal is for finding the questions you cannot answer while finding them is still free. It complements a human coach, never replaces one.

What’s next

This lesson closes the track’s first movement. Lessons 2 through 5 answered the first big question, how do I use this well: ask well, tailor the assistant, judge the handoff, run the method end to end. Lesson 6 opens the second question, the one this track was really built for: what does all this mean for my world? It starts by drawing the map: the kinds of risk this technology carries, and the levers for managing them. From here on you are not just using the machine. You are judging its place in your world.

If you remember one thing

Audience, platform, goal: three questions before every piece, one true story underneath them all.
Draft with the machine and rehearse against it, but keep every check for yourself, because the room is where you go alone.
AI can polish your story. It must never write your facts.