THE IDEA

Stay new. Sort the mess.

At 5:42 on Monday, Priya’s notebook says “Sam—badge,” “Jules?”, and the name of a tool she cannot spell. She does not need to sound senior by Tuesday. She needs tomorrow’s questions in one place.

Every cream box gives the model a job: Role, Task, Details, Format. The model drafts. You compare the draft with what happened.

Hard nos: No invented teammate, policy, decision, metric, or progress. No borrowed seniority.

Then do this: If a line did not come from your notes or a person who knows, delete it. Sounding new is allowed.

WHAT THE LABS FOUND

Notes first. Claims small.

Theo wants proof that AI will make him good at his new job. The research cannot promise that. It can show a few useful habits.

Hard nos: Do not turn a customer-support result into a promise for every job or new hire.

Then do this: Start with what you actually heard or did. Ask for a narrow shape. Check every claim.

BEFORE YOU PASTE

Use this script at your desk

June is about to paste her whole onboarding folder into a chat. It contains customer names, internal links, and a confidential document. She stops.

Replace full names with first names or roles when that is enough. Remove customer details, access links, passwords, private employee information, and documents you are not allowed to share. Mark uncertainty as [unknown] instead of guessing.

Hard nos: No confidential documents, secret links, passwords, restricted customer data, or private employee information.

Then do this: Read the material as if it appeared on a public screen. If that would make your stomach drop, remove it and ask a colleague instead.

HOW TO USE THIS

Four moves, then your review

Noah has twenty minutes before lunch. He chooses the unclear-task page, not every box in the booklet.

1. Pick the moment. Names, a message, notes, or one question.
2. Copy one box. Replace every bracket. Write [unknown] when needed.
3. Put notes before the final instruction. Keep the requested answer short.
4. Review against reality. Check every name, date, owner, deadline, decision, and claim of progress.

Hard nos: Do not paste material your employer forbids you to share. Do not let the model invent policy, approvals, reporting lines, metrics, or completed work. Do not send its wording untouched.

Then do this: Change at least one line into language you naturally use. If you cannot explain a sentence, do not send it.

DAY-ONE NAMES

Build a list, not an org chart

Priya’s notes contain three names, two arrows, and “ask Sam about badges.” She cannot remember whether Jules owns the tracker or merely showed it to her.

A neat table can quietly turn arrows into official reporting lines.

Then do this: Compare every row with the notebook. If a question mark became a fact, reply: “Questions stay questions. Do not infer relationships.”

THE GLOSSARY

Separate general meaning from local meaning

Luis hears “standup,” “ticket,” “SLA,” and “the board.” A general definition may help. It cannot tell him what those words mean on this team.

The likely failure is confidence: broad workplace knowledge gets presented as local policy.

Then do this: Star every “confirm here” item. Ask a human about the term needed for today’s work. Do not memorize the draft as company policy.

THE INTRO EMAIL

Warm, brief, and still new

June’s first generated draft calls her a “results-driven leader” and adds an achievement she never mentioned. Neither belongs in her introduction.

The failure is biography inflation. Smooth prose can smuggle in a career story you did not give.

Then do this: Read it aloud. Delete any claim you would feel awkward defending over coffee. Replace one polished phrase with words you actually say. You send it.

THE TASK I DON’T GET YET

Turn fog into questions

Noah is told to “clean up the tracker.” He does not know whether that means fix labels, remove old rows, or check owners. The model does not know either.

A plausible checklist may be completely wrong for this workplace.

Then do this: Ask the question that would most change your next action. If it directs edits, reply: “No changes yet. Clarify the assignment first.”

FIRST MEETING NOTES

Keep decisions separate from fragments

Imani leaves standup with “Sam blocked,” “Friday?”, and two acronyms. She wants a recap, not decisions nobody made.

The model may turn “Friday?” into “Due Friday” or attach an action to the nearest name.

Then do this: Highlight every action, owner, and date. If it was not explicit in your notes, move it to Questions.

THE HONEST STATUS

Say what moved—and what did not

Theo gets “Can you send status on this?” He opened the file and asked one question. A polished draft calls that “significant initial progress.”

The failure is padded progress. “Reviewed,” “advanced,” and “on track” may claim more than happened.

Then do this: Underline every verb. Ask, “Did I actually do this?” Replace any verb you cannot prove. Promise only a time you can keep.

THE FIRST 1:1

Bring three useful things

Claire completed one small task, still cannot find the badge process, and needs her manager to choose between two priorities. She needs an agenda, not a performance review.

The model may turn a first conversation into a speech about impact and growth.

Then do this: Say all three points aloud in under one minute. If they sound like a promotion case, reply: “This is week one. Use plain facts and one clear ask.”

WHAT I LEARNED

Make Friday useful, not shiny

Mateo learned where requests enter the team, mixed up two acronyms, and found the right person for access questions. That is enough for a real recap.

The danger is invented learning: a complete-sounding recap with conclusions you never reached.

Then do this: Point to the note supporting each bullet. If you cannot, delete it. Turn private frustration into a neutral question before sharing.

MONDAY FROM MY NOTES

Plan three hours, not a new life

Anya has two leftover questions, one promised update, and a 9:30 standup. She needs Monday morning to fit on one screen.

Generated plans often use every minute and add worthy-looking goals.

Then do this: Check the calendar and promised times yourself. Leave one open block. If there are more than three main actions, reply: “Protect the promise. Park the rest.”

THE SCARY QUESTION

Put the question where people can see it

Wes has wondered since Tuesday who owns the board. His draft begins with a paragraph apologizing for being new. The question is buried at the bottom.

The failure is avoidance disguised as polish. A kind paragraph can make the needed response unclear.

Then do this: Read only the first sentence. Can the person answer it? If not, reply: “Put the answerable question first.” Then send it or say it.

ONE SMALL THING A DAY

Match the box to the mess

June is tempted to run every prompt. She has one immediate problem: tomorrow’s 1:1. She opens that page and leaves the others alone.

Hard nos: Do not use ten polished outputs to avoid one human conversation. Do not treat the model’s plan as your manager’s priority list.

Then do this: Choose the page connected to the next real action. When that action is done, close the chat.

STUCK

Repair the reply, not your whole system

At 4:50, Luis gets a glossary full of confident company definitions. He does not need another app. He needs one sharper instruction.

Hard nos: Do not argue a false draft into being true. Do not keep a sentence merely because it sounds good.

Then do this: Check the corrected version against the source notes again. A second draft can introduce a second error.

FRIDAY DESK RESET

Leave a trail for next-week you

At 5:08, Imani has five tabs open and questions on three scraps of paper. Before leaving, she makes one small page for Monday.

Keep useful drafts beside the notes that support them. Follow your employer’s rules for tools, records, and company information.

Hard nos: No invented certainty. No borrowed seniority. No sending a draft you cannot explain.

Then do this: Put Monday’s first action where you will see it. Close the rest. Being new is not a defect to edit out.

SOURCES

Research behind page 4

Brynjolfsson, Erik; Danielle Li; Lindsey R. Raymond. “Generative AI at Work.” NBER Working Paper 31161, April 2023; published in The Quarterly Journal of Economics, 2025. Study setting: customer-support agents.
https://www.nber.org/papers/w31161

Chatterji, Aaron et al. “How People Use ChatGPT.” OpenAI, September 15, 2025; NBER Working Paper 34255.
https://openai.com/index/how-people-are-using-chatgpt/

Anthropic. “Be Clear and Direct.” Prompt-engineering documentation.
https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/be-clear-and-direct

Google Cloud. “Prompt Design Strategies.”
https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/prompts/prompt-design-strategies

Research can support a prompting habit. It cannot verify your employer’s people, policy, deadlines, or decisions.

Hard nos: Do not use a research result as proof that a generated workplace claim is true.

Then do this: Verify workplace facts with your notes, manager, or colleague.