Organizing & Processing Information with AI

Organizing and processing information

One of the most powerful things we can do with a large language model (LLM) is organize and process information. This is especially true for the kind of information that was very difficult for computers to work with before LLMs existed: notes, plans, messages, and lists that people wrote for other people. Working with that kind of information requires reasoning. A computer had no way to know which line of your notes was a date, which was a name, and which one mattered. AI can read the notes the way a person would and do the work we ask.

In this lesson we are going to learn three types of instructions for organizing and processing information. They are useful in almost every part of life, from studying, to organizing daily life, to a job.

• Extract: pull specific pieces of information out of something larger.
• Transform: change information from one form into another.
• Organize: arrange information in an order or in groups.

We will go through each one with examples, and then we will put them together.

Extract

To extract means to pull specific pieces of information out of something larger and leave the rest behind. We give AI the information and tell it what to pull out.

The first example is a recipe from a website. Recipe pages bury the recipe under a story. We want only the part we can cook from.

Look at what AI did. The text is one long paragraph with the recipe woven through the story. AI read every sentence, decided which pieces were ingredients and which were steps, and pulled out only those. It needed some reasoning to do it: “a good pinch of salt” became an ingredient, the aunt’s advice about “quiet” onions became “not browned,” and the sentence about the double batch was left behind. It did not add an ingredient the text does not mention. That is what extract does: the output is a piece of the input, chosen by our instruction.

The second example is an apartment lease. A lease is pages of clauses, and what most people need from it is a short list of when money is due and how much.

The email is one paragraph. AI pulled out nine dates and six amounts and labeled every one, including the recurring ones (“the 1st of each month”) and the relative one (“within 30 days of move-out”). Everything in the lists is in the email; nothing is invented. A person could now put the dates on a calendar and the amounts in a budget without reading the paragraph again.

The third example is a phone call. A mechanic explains what is wrong with the car, fast, and we are trying to write it down.

This time the instruction asked for two pieces per item, the work and the cost, and AI added the urgency because the mechanic said it for each one. It also separated what needs doing from what does not (the tires) and what was already done (the gas cap). When we write an extract instruction, we name the kind of pieces we want, and AI does the reading.

Transform

To transform means to change information from one form into another form without changing what it means. The pieces stay the same; the shape changes. We transform information so that it is consistent and usable: the same units, the same date format, the same kind of name.

The first example is a recipe again, but the problem is different. The recipe is in metric and the kitchen is not.

Every ingredient is still there, in the same order, with the same meaning. Only the units changed. AI knew that 115 g of butter is one stick, that 175°C rounds to the 350°F every American oven has a mark for, and that a 23 × 13 cm pan is the standard 9 × 5 loaf pan. Nothing was added and nothing was dropped. That is transform: the information in a form we can use.

The second example is a bank statement. The names on it are whatever the card machine sent.

AI transformed eight strings that only a computer could love into names a person recognizes. It knew that “SQ ” and “TST” are payment processors, not businesses, and that the numbers after Shell and Kroger are store numbers. This is the transform everyone does by hand when they read a statement and think “what was that?”

The third example is contact information scattered through a message thread. We want it in one table.

Four messages written four ways became four rows in the same four columns. Phone numbers written with dots, parentheses, dashes, and nothing at all are now one format. Notice that this one instruction did an extract and a transform together: AI pulled the four pieces out of each message and put them into one shape. That is common. The instructions are separate ideas, and we often use them together.

Organize

To organize means to arrange information in an order or in groups. Extract chooses the pieces and transform gives them a common shape. Organize decides where each piece goes, so that the arrangement answers a question.

The first example is a pile of receipts from a weekend trip. In time order, they become a record of the trip.

That is the whole weekend, reconstructed from a pocketful of receipts, in the order it happened. AI read nine dates written five different ways, put them in order, and grouped them by day. If someone asks “what did you do in Chattanooga?”, this answers it, and the receipts never could.

The second example is a to-do list, the kind that piles up in a notes app with dates written any which way.

Ten items in no order became three groups in date order. AI worked out that “friday” means October 9 from the date we gave it, and that “the 17th” means October 17. Nothing was added; the three undated items stayed undated instead of being given made-up deadlines. This is the list to look at on Monday morning.

The third example is a list that needs grouping by something AI has to know.

The list had no years in it. AI supplied them from what it knows, then organized by the rule we gave. When AI adds information like that, it is worth checking, but the organizing itself is exactly what we asked for. The same organization we choose decides what we see: the to-do list by date shows what is urgent; the same list grouped by where the errands are would show a route.

Putting it all together: an outline

Think about the table of contents in a book. It is a list of headings, and each heading describes what is in one part of the book. Each heading was extracted from its part: it names the content and ideas found there. All the headings were transformed into a common format: short, the same style, the same length. And they were organized into an outline: chapters in order, sections under chapters. A table of contents is all three instructions applied to a book.

We can do the same thing to notes. Here are notes typed during the first day of a job at a coffee shop, straight from whatever the manager said, in the order she said it.

Look at the outline and find the three instructions in it. AI extracted every fact from the notes: fifteen facts written in shorthand became fifteen bullets, and nothing was dropped. It transformed each fact into a common format: a short phrase, capitalized the same way, with the abbreviations spelled out (“sunday nights” became “Sunday nights,” “5 min grace” became “5-minute grace period”). And it organized the bullets into groups, wrote a heading that describes each group, and put the groups in a sensible order, with the schedule first and parking last. The notes and the outline contain the same information. The outline is the one we can use at 5:45 on a Monday morning.

A bullet can be broken down further

Any bullet in an outline can be treated as a piece of information in its own right and broken into smaller bullets. Here is the same job outline, and then a request to expand one bullet. Compare the two replies.

The outline is the same except for one bullet, which now has six steps under it. The steps did not come from the notes. The notes only said that backflushing happens at close. When we ask AI to break a bullet down, the new information has to come from somewhere, and there are three possibilities.

• From something we provided. If the red binder had been pasted into the conversation, the steps would come from it. This is extract again, and it is as reliable as our source.
• From what AI already knows. AI has read a great deal about espresso machines, so it filled in the usual procedure and told us that is what it did. This is often right and always worth checking.
• From invention. If nothing in the conversation or in training covers a bullet, AI will still produce sub-bullets by making them up.

Each of the three can be useful. When we want the facts, we want the first and we should supply the source. When we want a reasonable starting point, the second is fine, and we check it. When we want ideas, such as six possible names for a new drink, invention is exactly the job. The skill is knowing which one we are getting, and asking for the one the job needs.

Applying extract, transform, and organize to achieve different goals

The three instructions are not only for notes. Any time information arrives in a form that is hard to use, we can extract the pieces that matter, transform them into a common form, and organize them for the question we have. Here are five situations where that turns something useless into something useful.

The first is a credit card statement. It is a list of cryptic merchant names and amounts, in date order, which answers no question anyone actually has.

Find the three instructions. AI extracted the merchant and the amount from each line. It transformed each merchant name into a category: “SHELL OIL 4218” became gas, “KROGER” became groceries, “STEAM PURCHASE” became entertainment. It knew what those merchants sell. Then it organized the charges by category, added them up, and put the biggest category first. The list of subscriptions is a second extract on the same charges, using a different rule. Twenty lines that meant nothing became a page that says where the money went.

The second situation is cooking. Three recipes have three ingredient lists, in mixed units, with the same ingredients repeated. What we need at the store is one list.

AI extracted every ingredient from the three recipes. It transformed the amounts into common units so they could be combined: “16 oz chicken breast” became 1 lb, and three separate amounts of garlic became 9 cloves, which it also translated into something you can buy, one head. It kept the two kinds of chicken separate, because they are different things at the store. Then it organized the list by aisle, so the trip through the store goes in order. The last group has a note, because oils and soy sauce are things most kitchens already have. Three recipes became one list you can hold in one hand.

The third situation is sports. A game is a stream of events. A stat sheet is those events extracted, transformed into common statistics, and organized by player.

AI extracted the events: each goal, each assist, each card. It transformed each one into a statistic, so “Reyes scores, pass from Kim” became one goal for Reyes and one assist for Kim. Then it organized them twice. By player, the table lets us compare people. By half, the table lets us see when the game turned. The same eight events produced two tables, because we asked for two organizations. Every organization makes something visible and hides something else. Choosing the organization is where our thinking goes.

The fourth is a work trip. The receipts are the same kind of pile as the Chattanooga weekend, but now someone has to be paid back, and the person approving it wants a specific form.

This is all three instructions in one prompt. AI extracted the date, vendor, and amount from each line, transformed every line into the same row shape and every date into one format, and organized the rows by date with totals by category. It also did the part a person doing this by hand would do: it split the minibar out of the hotel bill and flagged the drugstore receipt. The report is ready to submit, and every number in it can be checked against a receipt.

The fifth is an email thread. Twelve messages about one event, and the decisions are buried in the replies.

Twelve emails became six decisions and seven tasks. AI extracted the decisions from replies like “Community center it is” and “$25 is fine,” where the decision is a sentence fragment, and it extracted tasks from offers like “I can call them,” where the task was only implied until Dana confirmed it. It transformed each one into the same shape: a task, an owner, a deadline. And it organized them into two lists, with the tasks in a table because a table is the form a to-do list wants. Anyone who missed the thread can read this in thirty seconds and know what they owe.

What you should be able to do

By the end of this lesson, you should be able to do three things with any information that arrives in a hard-to-use form.

1. Write an extract instruction. Name the kind of pieces you want (ingredients, dates, amounts, repairs, contacts, decisions) and let AI pull them out of the whole.
2. Write a transform instruction. Name the form you want every piece in (one unit system, a plain name, one phone format, a short bullet, a row in a table), and check what the transform lost.
3. Write an organize instruction. Name the order or the groups you want, and try a second organization when the first one does not answer your question.

Often one instruction does all three at once, as when we ask for an outline, a spending analysis, or a shopping list. When you write that instruction, you are still deciding what to extract, what form to transform it into, and how to organize it. AI does the reading and the arranging. You decide what matters, what shape it should take, and whether what came back is true.