Take the Quiz
Shift Library
Each issue starts with a real work moment and leaves behind one useful move: a prompt, a check, or a sharper way to think about it.
The Shift Library
Read any one in a couple of minutes. Each ends with something you can use before your next meeting, message, or decision.
When a client email asks several things and your answers are scattered through your notes, give AI the email and the notes as two blocks. Ask for one numbered answer per question, from your notes only, with anything still open listed for you.
Project notes use words that depend on each other: a passed pilot only lets you ask for acceptance. Give AI one line per term, what it means and what it does and doesn't lead to, before it writes the client update.
Before a weekly check-in, type your answer to the follow-up you dread and have AI check it against your own update notes for the one line that would mislead. Then try again in your own words.
When two reviewers ask for edits that can't both fit in one email, a single rewrite has already picked a winner. Ask AI for the finished email both ways, each naming what it leaves out, and make the call yourself.
A polished AI-assisted draft can look equally checked everywhere. Send it with a short note that separates what AI did from what you checked, names who decides and asks your reviewer one question only they can answer.
Marco's invoice came in $1,860 over his quote, and $600 of that is a page you approved by email. Sort every line against the quote and your written approvals before you ask him anything.
Before you present an AI-assisted spreadsheet analysis, close the chat and answer three questions about one output: what it represents, which inputs determine it, and which direction it should move if one input changes.
An AI follow-up list that finds no completion evidence might be accurate and still wrong. Add a 'status unverified' column so the list tells you what it checked before you send a nudge.
Your chart and your title agree. That's not the test. Strip the title and ask AI what the data says on its own, because your audience reads the shape before the words.
When you combine two spreadsheet exports with different row structures, values repeat on every matched row and the total inflates without a single cell looking wrong. Ask AI for a merge preview before you trust the combined number.
When software gets confusing, a screenshot gives AI the context your words leave out. But only if you say what you're trying to do.
OpenAI's Sponsored Agents label conversations inside ChatGPT. The label disappears when you paste the answer into your own vendor comparison. Add a source column before that happens.
Every paste into AI is a data decision most people make on autopilot. Answer three questions before you paste: what kind of data, can the tool act or only read, and does someone else need to know.
AI translations are fluent, but fluent is not faithful. Before you send a translated client email, check whether the commitment level survived, not just the grammar.
Every AI case study leads with hours saved. Track how often you correct, rewrite, or override what the AI produced. That tells you whether it's working.
Your notes mix facts, guesses, and copied text. Before AI summarizes them, label each line so the model doesn't flatten everything into one confident voice.
AI meeting prep gives you talking points that confirm what you already think. The sharper prep is the one question whose answer would reverse your recommendation.
When you paste a vendor PDF, a client email, and your own notes into the same prompt, AI treats all of it as equally reliable. A three-line header changes that in thirty seconds.
Your draft reads clean, but polished language can hide thin reasoning. Ask AI to tag each sentence as explanation, assertion, evidence, or unsupported claim. The unsupported ones are where your reader will push back first.
When AI can reach a large library of internal documents, a stale file can slip into the answer. Name the source pack, set a freshness window, and check what the model used.
When an AI draft reads well but the reasoning is thin, re-prompting won't fix it. Four questions will: claim, evidence, assumption, counterpoint.
A forty-message AI thread is a transcript, not a handoff. Before you close a working session, ask AI for a five-line restart brief: goal, decisions, open questions, sources, and next action. Tomorrow starts in thirty seconds.
Most AI meeting summaries are organized around what happened. Turn your notes into a handoff card organized around what happens next.
If your first agent writes a confident draft when it should stop, every agent you connect to it inherits the problem. Write a single-agent brief before you scale.
AI can give you ten polished reasons without proving that any of them are strong. Give it the full situation, test the strongest reason and ask what evidence would change the answer.
Most AI meeting prep collapses two opposite jobs into one prompt. Use AI to expand your thinking before the call, then compress to decisions and owners after it.
Most model debates aren't about the model. They're about a missing work standard. Answer three questions before you pick a tool, and the argument disappears.
Most AI drafts make your case sound more confident. One three-question prompt catches the gap between polished and right, in thirty seconds, before the client catches it for you.
Most teams aim AI at faster work inside one role. The bigger win is at the handoff where one person's output becomes someone else's input. A cross-role brief translator prompt handles the translation.
AI is already in your email, your project tool, and your coworker's deck. The useful question isn't whether to adopt it. It's which parts of your work it genuinely improves and which stay yours.
Don't ask AI if your email is good. Make it read the message as the person getting it, and name the line that lands wrong.
AI made raw horsepower nearly free. The same flood reprices the one thing it can't make, your judgment, upward.
"Make it more compelling" changes nothing. When AI output is flat, add nouns, not adjectives.
The slow part of merging inputs isn't the merging, it's translating five formats by hand. Hand over the mess unedited and let one prompt flag the clashes.
A sharper dictation model just landed in ChatGPT. The real win is what it pulls out of you.
Two new models just launched. The move that pays off isn't picking one, it's aiming the AI you have at the task you repeat.
Every third post says someone automated their whole job. You're grading your rough cut against their final cut.
AI answers with only what you gave it, and sounds just as sure either way. Ask what it's missing.
Stop rebuilding the same prompt every week. Package the task into a tool, and build it to interview you for the inputs.
The dangerous AI errors aren't the fabrications you catch. They're the plausible ones, wrong only on the clause your decision rests on.
A custom GPT is a worker you build once and hand out. A Project is a workspace where one area's context accumulates. Match the tool to what you're reusing.
A dense requirements doc is a perfect spec for AI. Hand it over whole with a prompt that drafts against every requirement and shows coverage.
You've briefed AI the same way forty times. Here's how to make it stick, once.
AI doesn't know your bar. Skip the rubric you'll never write, show it five calls you've already made, and it picks up your standard.
"Is this plan good?" gets you flattery. "Why did it fail?" gets you the risks.
A talk loses the room because it's predictable, not underprepared. Open on the stakes and keep a gap open the whole way.
A generic answer isn't the model's ceiling. It's the ceiling of what you asked. Hand AI a job, not a question.
The win isn't the first output. It's the third, after you push back and your standard becomes visible.
A speed win can quietly remove the check, not just the effort. The slow step was where problems got caught.
AI raises the ceiling on how much you take on. The real skill is refusing to fill the space it clears.
When a project's next steps live only in your memory, you are the single point of failure. Give AI the standing context.
AI can build the work. The part you most want to offload, checking it, is the part they pay you for.
The ability you fear AI is dulling was never your durable edge. Pattern recognition is, and it compounds.
Turn a vague AI automation idea into a clear request with trigger, input, output, owner, approval path, and test case.
A practical AI meeting prep path for finding the decision, risk, and one useful question before the call starts.
Use AI to find the real decision, missing information, and next reply inside a messy work thread.
Next step
