# The prompts I actually use

These aren't clever prompts. There's no magic phrasing. They're the plain sentences I say over and over - I dictate more than I type - because each one does a specific job: keep the AI honest, keep it anchored to what I actually know, and keep me in charge of the work.

This is the base level, on purpose. There are more advanced moves for cross-checking sources against files - that's its own page, another day. These are the ones I use every single session.

Copy the ones that fit how you work. Change the wording. The point isn't the exact words - it's the habit each one builds.

## To catch fabrication

The most important category. AI will state something wrong with total confidence, so these are the questions I use to poke at any answer that matters.

- **"Is that true?"** - The bluntest one. Surprisingly often, this alone makes the AI walk back a claim.
- **"Did you make this up?"** - For anything that sounds too clean or too convenient.
- **"Show me a URL where I can open the source and read it for myself."** - Not "cite your sources." A link I can click, open, and check with my own eyes.
- **"Did you actually search for the real-time answer, or did you guess from your memory?"** - I ask this constantly. An AI's memory has a cutoff and a confident tone; those are a bad combination. This question makes it tell you which one you're getting.

Then the part no prompt can do for you: open the source and read it. The checking is yours. It's still not foolproof - but I'm a lot better at it than I used to be, and these four questions are most of the reason.

## To keep it anchored to what I know

An AI defaults to a generic version of the world. These pull it back to your actual context - your notes, your files, your situation - before it answers.

- **"Using what I've given you, how should I approach this?"** - Anchors the answer to your material, not the internet's average.
- **"Did you read the file I gave you, or are you guessing?"** - A legitimate check, not a rude one. Ask it any time an answer sounds generic.
- **"Read the entire file. Every page - not the title, not the abstract, not the first page and a guess about the rest."** - This one comes from experience. In my early sessions I fed in journal articles and discovered the AI had read the filename and improvised the rest. Skimming is what these tools do by default when a file is long. Say the whole sentence.
- **"When you're done, tell me honestly: did you read all of it, or did you skip? If you skipped, say what."** - The companion move, and it works better than it has any right to: an AI will quietly skim, but when asked directly, it will usually admit a partial read. Make the honesty part of the job.
- **Spot check: "Quote me the last paragraph of the file."** - A skimmer fails this in one line.

## To keep the work on point

Factually correct is not the same as relevant. A paragraph can be true, well-written, scientifically solid - and useless for the piece you're actually making. These are the questions for that.

- **"Why are we doing this - and is what you just made still serving that?"** - The purpose check. Run it whenever the work starts feeling thorough instead of useful.
- **"Who is this for, and is this how they actually work?"** - Correct-in-the-field is not the same as true for your reader. The AI knows the textbook version; you know the audience.
- **"Every piece has a point. Is this piece still keeping its point?"** - For work that lives among other pieces - posts in a strategy, chapters in a doc, pages on a site. Each one has a defined job; drift is silent.
- **"Is the story advancing, or did we just add another paragraph that's true but useless?"** - For one big document. Information that doesn't move the story is padding with citations.

## To stay in charge of big work

These put a checkpoint in front of the AI before it runs off and does a lot of work in a direction you didn't want.

- **"We're just planning here. Don't do anything yet."** - The mode-setter. Brainstorm, weigh options, think it through together - and nothing gets built or changed until you say so.
- **"Show me the outline first. Let's look at it together before you write."** - Never let it produce the whole thing before you've seen the plan.
- **"Do the first step only, then stop and show me."** - Breaks a big task into checkable pieces.
- **"Audit this independently. Spin up a separate agent, write its checking prompt, and have it review this work."** - My favorite advanced-but-easy move: I don't interrogate the result myself line by line - I make the AI build its own auditor. A fresh set of eyes that didn't do the work finds what the doer can't.

## To manage the session

- **"Show me the big picture, start to finish - a file with checkboxes and status, and keep it updated as we go."** - I go down rabbit holes. They're often useful and always distracting, and the status file is how I find my way back. (A plain checklist file beats a fancy dashboard; fancy breaks.)
- **"Where are we, and what's the next step?"** - The re-orient. I need to see the next step and where it sits in the whole, every time I surface.
- **"Is this saved, or does it only exist in this chat?"** - I ask this a lot. Decisions and finished work go into files; a chat scrolls away, a file you can come back to.
- **"Have I missed anything? What gaps didn't we even think to consider?"** - Because I don't know everything, and neither does the session. For big work, combine it with the auditor move above: an independent set of eyes, briefed to look for what's absent.
- **"Explain what we're doing to someone who wasn't in this conversation."** - A quiet test of whether you actually understand the work, or you're just borrowing the AI's understanding. If you can't explain it, you don't have it yet.
- **"Let's write up where we got to before we stop."** - End on purpose. Get the state into a file so tomorrow doesn't start from zero.

## The pattern underneath all of them

Notice what these have in common. None of them ask the AI to be more impressive. They ask it to show its work, to read what you gave it - all of it - to stay on your ground, to stop and check in, and to prove it's telling the truth. That's the whole posture: the AI does the busywork, you stay the one in charge.

If I can do it, you can do it. Pick three of these, use them this week, and watch how much less often you get quietly handed something wrong.