7 Ways to Use AI Without Generating Text

Using AI without generating text means the model never writes a word you keep. It finds, it ranks, it matches, it compares, it points at what to delete. Everything that comes out the other side was already yours. Editor and librarian, never author.

Here's the part that gets missed. Generation is the loudest capability these tools have, so it swallowed the whole conversation, and meanwhile the thing most people are drowning in isn't a shortage of words. It's a pile of words they can't find. McKinsey Global Institute measured the search tax and put it at 9.3 hours a week per employee, close to a fifth of the workweek, spent hunting for internal information. That was 2012, back when you had a fraction of the files you have now. Writing more was never the bottleneck.

Seven ideas below, and none of them asks a model to compose anything. If the ethics of generated work are what brought you here, the sibling piece on using AI without stealing anyone's creative work covers that ground directly. This one is about a quieter question: what's it good for when you take authorship off the table? Each idea ends with something you can do this afternoon.

1. Find Out Whether You Already Wrote It

Before you ask a model to draft anything, ask it whether a version already exists in your own files. Most of the time one does. You wrote it in a proposal in March, or buried it in the third paragraph of a message to a client, and you remember the idea perfectly while remembering none of the words that would find it.

That gap is why keyword search keeps failing you. Your file search matches strings. Your memory stores meaning. Semantic search closes the gap: it matches on what a passage is about rather than which characters it contains, so "the thing I wrote about why we stopped doing fixed-price work" lands on a document that never used those words. NotebookLM is the easiest way in, because it only answers from sources you give it and cites the page it pulled from, so you can go read the original instead of trusting a paraphrase.

The reframe is small and it changes your whole week. You're not short of writing. You're short of retrieval.

Start today: Pick the one folder where your best thinking goes to die. Upload it, then ask a full-sentence question you'd ask a colleague, not a keyword: "where have I explained how we price a rush job?" Read what comes back before you write anything new.

2. Let It Choose Which Testimonial Goes on Which Page

Renata runs a two-person landscape design studio outside Asheville. She has 61 reviews spread across three platforms and uses the same three on every page of her site, because picking between 61 of anything is miserable and those three were the first ones she copied over two years ago.

Here's what she's leaving on the table. The best testimonial for a page is almost never the most enthusiastic one. It's the one that answers the specific worry the person reading that page is carrying. Someone on a pricing page is holding one question: will these people blow through my budget. Someone on a portfolio page is holding a different one: will this look like my yard or like their portfolio. A five-star review that says "wonderful to work with" answers neither.

So hand a model all 61 quotes and one line about the objection that page has to beat, and ask it to rank them. Quoted exactly, no editing, no smoothing. The output is a shortlist of other people's real sentences, chosen better than you'd choose them at eleven at night.

Start today: Export every review, testimonial, and unsolicited thank-you message you've ever received into one document. Name the single objection your most important page has to beat. Ask for the five quotes that answer it, verbatim, with one line on why each one works. Paste the winner in and change nothing about it.

3. Find the Thing You Keep Circling

You've been making the same argument for four years and you can't see it, because you made it one piece at a time, months apart, under different titles. Everyone who reads all of your work at once can see it. Nobody reads all of your work at once, including you.

A model does, and clustering is the job it's quietly excellent at. Not summarizing the pile. Grouping it, by the claim underneath rather than the topic on top, and telling you which group is biggest. Two posts about hiring and one about a software migration can all be the same argument about what you refuse to outsource, and that's the kind of pattern that stays invisible from the inside.

What you find is usually your positioning, sitting there fully formed and never stated. It's also the map for turning a back catalog into something sequenced, which is the starting point for turning one piece into a week of material.

Start today: Paste the title and opening paragraph of everything you published in the last year into one document. Ask for groups by underlying claim, not by subject, with the titles listed under each group and no commentary. Look at the biggest group. That's what you're really doing.

4. Build a Week of Posts Out of Sentences You Already Published

Repurposing has a bad reputation and it earned it honestly. You feed in a good article, a model hands back five posts that sound like a stranger doing an impression of you, and you delete them. The problem isn't repurposing. It's the paraphrase step in the middle.

Take the paraphrase out and the whole thing works. One rule: zero new words. Every post is built only from sentences that appear in the original, in their original wording, selected and ordered. The model's job is to find the eight lines that stand alone without the paragraph around them, which is a genuinely hard selection problem and one it's good at. Your job is to pick which ones ship.

This is the item you stop doing by hand. Once you've defined the rule once, running it against every new piece is mechanical, and that's exactly the shape of thing you hand off to a workflow that reads your library and returns quotable lines every time you publish.

Start today: Take the best thing you published last month. Ask for the ten sentences that survive on their own, quoted exactly, with nothing added, joined, or tidied. Check three of them against the original word for word. Then schedule the four you like.

5. Find Everything You Need to Remove Before You Share It

Nadia spent a weekend building the case study that would anchor her freelance UX portfolio. Twenty pages, real research, the best work she'd done. Her client's name appears in the deck eleven times: in two screenshots, in a participant quote, in a file path visible in the corner of a screen recording, and once as an axis label on a chart she rebuilt from their dashboard. She found seven of the eleven.

Finding is the hard part, not deciding. You know perfectly well what has to come out. What defeats you is that it's spelled four different ways, abbreviated in one place, and sitting inside an image in another. A model reading the whole document at once catches the ones your eye slides past, and hands you a list of locations rather than a cleaned file, which is the version you want. You do the deleting.

This pattern has an industrial-strength cousin worth knowing about. Microsoft's Presidio is a free open-source toolkit built for exactly this, detecting names, phone numbers, account identifiers, and addresses across documents at volume. For one deck you don't need it. For every deck that leaves your company, it's the sort of check that belongs in the step that runs before a document goes out.

Start today: Take the next thing you're sending outside your organization. Ask for every place a person, client, price, internal system, or unreleased name appears, as a list with locations and nothing else. Work down the list by hand. Count how many you'd have missed.

6. Ask for Cuts, Never Replacements

Your voice dies in the rewrite. Not in the idea, not in the structure, in the moment a model takes your sentence and hands back a smoother one. Smoother is the problem. Your rhythm is the thing people recognize, and smoothing is precisely what removes it.

So ban the rewrite and ask only for subtraction. Which sentences can come out without losing anything? Which paragraph is making the same point as the one above it? Where does this say the same thing twice in different clothes? Everything that comes back is a list of your own sentences with a verdict next to each, and you can disagree with every single one at no cost.

Honest note: a decent share of this needs no AI at all. Hemingway Editor has highlighted overlong sentences and passive constructions for years without writing a syllable, and reading your draft out loud beats both. Use the model for the judgment call it's better at, which is spotting the two paragraphs that are secretly the same paragraph.

Start today: Take a draft you've been fiddling with. Ask for cut candidates only, each with a one-line reason, and say explicitly that no replacement text is wanted. Accept none of them automatically. Cut two.

7. Find Where the Outdated Fact Still Lives

You changed your pricing in April. You updated the pricing page the same day. Nine months later a prospect quotes your old number back at you, having found it in a PDF, a help article, and the third slide of a deck you still send.

Facts about you leak into more places than you can hold in your head, and they never leak in the same wording twice. That's the whole reason this is a job for a model rather than find-and-replace. Your old price appears as a number, as a range, as "under two thousand," as an example in a case study. Search for the string and you'll find one of the four.

Be honest about the split, though. If the thing you're hunting is an exact string, a plain text search is free, instant, and more reliable than any model. ripgrep will tear through a folder of files in under a second and never hallucinate a match. Reach for AI when the fact is phrased differently every time it appears, which is the case that breaks the free tool.

Start today: Name one fact about your work that changed in the last year. Price, title, team size, a service you dropped. List the four places your words live. Ask for every passage that contradicts the current truth, with the location of each, and fix them in one sitting.

Pick One

Seven ideas, and the honest prediction is that trying all seven this week ends with none of them done. Take the one that made you stop reading for a second. If none did, take number one, because ten minutes of searching your own archive is the cheapest of the seven and it tends to embarrass you in a useful way.

There's a bigger idea hiding in all of this. The work you've already done is a much larger asset than the work you're about to do, and almost nobody treats it that way. If you're new to this site: the Sunday column covers the week's AI news, the Wednesday one takes a single workflow apart and rebuilds it, and Monday is for ideas like these. When one of them stops being an idea and turns into a chore you repeat, the workflow directory shows what other people handed off.

Frequently Asked Questions

• If the AI isn't writing anything, what is it doing?+
  Four jobs: finding, choosing, comparing, and removing. Retrieval pulls the thing you already made out of a pile you can't search by keyword. Selection ranks what you own against a goal. Comparison tells you where two versions of a fact disagree. Removal nominates what to cut. All four operate on material that already exists, so nothing it hands back is invented. If you want these running unattended rather than by hand, the overnight edition of this column covers how that works.
• Will the output still sound like me?+
  It will, because nothing gets rewritten. That's the whole point of locking the instruction to verbatim extraction. The failure mode to watch for is silent paraphrase: a model will tidy a sentence and hand it back as a quote. Ask for the location of every quote alongside the quote itself, then spot-check three against the original. If a word moved, say so and ask again.
• What do I need to set this up?+
  Whatever chat tool you already use, plus a way to hand it your own material. For a folder of documents that means either uploading them or connecting the drive they already live in. Free tiers handle every idea here at small scale. The limit you hit first is how much you can paste at once, which is what pushes most people toward a connected folder.
• Isn't this still AI trained on other people's writing?+
  Yes. The model is the same model, and using it this way doesn't change how it was built. What changes is what comes out the other side: your own sentences, your own quotes, your own files. If the provenance of the model itself is the thing bothering you, that's a separate and legitimate question, and it deserves a real answer rather than a workaround.