Transform and Analyze: Summarize, Extract, Classify, Rewrite, Translate, Compare

Language models are at their best when they work on text that you provide. Summarizing a report, pulling the key facts out of an email, sorting feedback into categories, rewriting for another audience, translating, comparing two versions: these everyday jobs are where AI saves the most time, and where a few good habits make results much more reliable.

You now know how to structure prompts and shape outputs. This lesson applies those skills to six families of transformation tasks. For each one you will see what to specify and what to watch out for. These are the building blocks of most prompts in your future library.

A general principle: bring the material

When the model transforms a text you gave it, it is anchored to a real source, so hallucinations are less likely than when it answers from memory. Still, it can omit, distort or add. In every task below, the safest formula is:

1. Provide the text clearly delimited.
2. Say precisely what to do and for whom.
3. Tell it to stay inside the text.
4. Ask for a format that makes checking easy.

1. Summarize

"Summarize this" is too vague: a summary for a busy director differs from one for a student. Specify the purpose, the reader, the length and what must be kept.

Summarize the report below for a manager who has two minutes.
Format: a one-sentence conclusion, then 4 bullet points (what
happened, what it means, risks, what to decide).
Keep all figures exactly as written. Do not add information.
If something important seems missing, mention it at the end.

Useful variations: an executive summary, a "TL;DR," a summary by section, a summary for a specific question ("What does this contract say about cancellation?"). For long documents, summarize in parts first, then summarize the summaries, and spot-check against the original.

2. Extract

Extraction pulls specific items from a text: names, dates, amounts, decisions, tasks, questions. It is a precise task, so be explicit about the fields and the behavior for missing data.

From the email below, extract: sender, request, deadline, and any
amount mentioned. Present them as a short list.
Rules: copy the wording from the email; write "not mentioned" if
a piece of information is absent; do not infer anything.

Asking for a supporting quote next to each extracted item makes checking fast. If you plan to reuse the data (in a spreadsheet, for example), request a table or structured format, as seen earlier in this chapter.

3. Classify

Classification sorts items into categories you define: customer comments by theme, emails by urgency, ideas by feasibility. To get consistent results:

• Define the categories, each with a one-line description.
• Include an "other / unclear" category so the model does not force-fit items.
• Give one or two examples for borderline cases.
• Ask for a short justification for each decision.

Classify each comment as: Delivery, Product quality, Price,
Support, or Other.
Definitions: Delivery = anything about shipping time or damage in
transit. Support = anything about contacting the company.
Output a table: Comment number, Category, One-line reason.
If a comment fits two categories, choose the main one and note it.

Be careful if the classification affects people (rating applicants, judging performance, grading): bias and errors matter much more there, and you should keep a human in charge of the decision.

4. Rewrite

Rewriting covers proofreading, simplifying, shortening, changing the tone or adapting to an audience. The key is to define what must not change.

Rewrite the paragraph below for a general audience with no
specialist knowledge.
Keep: all facts and numbers. Change: vocabulary and sentence length.
Maximum 100 words. Then list any term you simplified so I can check
that the meaning is preserved.

For proofreading, ask the model to list its corrections ("show each change as before and after") so you stay in control of your own style. For "make it sound more professional," specify what that means to you, or you will get clichés.

5. Translate

Models translate well between many languages, but quality varies by language pair and by text type. To get a useful translation, give context:

Translate the text below from French into English.
Context: it is a friendly newsletter for members of a local
association. Tone: warm and natural, not literal.
Keep proper names and dates unchanged. If an expression has no direct
equivalent, give your best natural option and flag it in a note.

• Specify the audience and the register (formal "you" versus informal, regional variants).
• Ask for notes on ambiguous passages.
• For legal, medical, technical or official documents, treat the result as a draft and have it checked by someone competent in the language and the field.
• A useful check: ask the model to translate the result back, then compare with the original meaning.

6. Compare

Comparison is excellent for two versions of a text, two offers, two options in a decision. Name the criteria and the format.

Compare version A and version B of this policy paragraph.
Output: a table with the columns Point, Version A, Version B,
Impact for staff. Then list changes in meaning, not just wording.
Quote the exact passages where versions differ.

Version A: [text]
Version B: [text]

When comparing options for a decision (for example, two suppliers), provide the facts, ask the model to use only those, and treat its recommendation as an input to your own judgment, not as the decision.

Chaining transformations

These tasks combine naturally. A typical chain for feedback analysis: extract the comments from a document, classify them by theme, summarize each theme, then rewrite the result as a short note for your team. Remember the human checkpoint after extraction.

Common pitfalls

• The model "improves" content you wanted untouched: say explicitly what must stay identical.
• It fills gaps: always provide a rule for missing information.
• Inconsistent categories: define them, add examples and an "other" option.
• Sensitive content: summaries and extractions of documents with personal data create exactly the privacy risks you will study in chapter 4. Anonymize first.

Your turn: transform a real text

Exercise 1. Take a text of at least 300 words that you may share (a public article, your own notes). Write one prompt that asks for a two-part result: a three-bullet summary, then a list of every figure or date mentioned, each with the sentence it comes from.

Exercise 2 (project). Choose one of these for your library: a "summarize for a manager" prompt, an "extract action items" prompt, or a "rewrite for a different audience" prompt. Define the variable parts, add a rule for missing information and a built-in way to check the result.

Use only the text below.
Part 1: Summarize it in exactly three bullet points of at most
20 words each.
Part 2: List every figure and every date it mentions. For each,
quote the sentence it comes from, in quotation marks.
If the text contains no figures or dates, write "none".

--- TEXT ---
[paste the text here]
--- END ---