You just got a 40-page vendor proposal at 4:55 PM. Your boss wants the gist by morning. Instead of reading until midnight, you paste it into an AI tool and ask: "Give me the five key takeaways, the top three risks, and a one-line summary I can forward to my manager."
Ninety seconds later, you have a draft. You skim, sanity-check the numbers against the document, tweak the wording, and send it. That is the core skill in this lesson: turning hours of reading into minutes of review.
Large language models (the technology behind ChatGPT, Claude, and Gemini) are pattern machines trained on enormous amounts of text. "Summarize," "compare," and "extract" are exactly the kinds of language tasks they handle well, because they have seen millions of reports, contracts, and emails.
Three everyday jobs they do reliably:
The catch: AI can be confidently wrong. It can invent a number or miss a clause. So the rule is simple. AI does the first pass. You do the final check. Always verify anything that drives a decision.
Weak prompt:
> Summarize this report.
Strong prompt:
> You are helping me brief a busy executive. Summarize the attached report in three parts: (1) five key takeaways as bullets, (2) the top three risks, (3) a one-sentence summary under 25 words. Keep it plain and specific. Quote exact figures where they appear.
In 2026, all three major tools accept files directly. In ChatGPT, Claude, or Gemini, you click the paperclip or upload icon and attach a PDF, Word doc, or spreadsheet. For very long documents, Claude and Gemini handle big files especially well thanks to their large "context windows" (the amount of text a model can read at once).
If you only have plain text, just paste it in.
Ask the AI to cite where each point came from:
> For each takeaway, add the page or section it came from so I can verify it.
This makes spot-checking fast and discourages the model from making things up.
Sometimes you do not want a summary. You want one number buried on page 22.
Real scenarios:
Example prompt:
> From this contract, extract every payment obligation. Return a table with three columns: Amount, What it pays for, Due date. If a due date is not stated, write "not specified."
That last instruction matters. Telling the AI what to do when information is missing stops it from guessing.
For a free, well-explained primer on writing clear instructions, see Google's Prompting Guide, which applies to any tool, not just Gemini.
This is where AI saves the most time. Say you have two proposals, "Vendor A" and "Vendor B," and you need to recommend one.
Upload both documents, then prompt:
> I am choosing between these two vendor proposals. Build a comparison table with these rows: Annual price, Setup fee, Contract length, Support hours, Data security certifications, Cancellation terms, and Notable risks. Add a final row with your recommendation and a one-line reason. If a proposal does not mention something, write "not stated."
You might get back something like this:
| Criteria | Vendor A | Vendor B |
|---|---|---|
| Annual price | $48,000 | $39,500 |
| Setup fee | $5,000 | $0 |
| Contract length | 12 months | 24 months |
| Support hours | 9-5 weekdays | 24/7 |
| Security certs | SOC 2 | SOC 2, ISO 27001 |
| Cancellation | 90-day notice | No early exit |
| Notable risks | Higher cost | Locked in 2 years |
What took two hours of cross-referencing now takes two minutes. But notice: Vendor B is cheaper *and* offers better support, yet locks you in for two years with no exit. That tradeoff is the real decision. The AI surfaced it. You make the call.
Once the table exists, keep going in the same chat:
> Which vendor is better if my top priority is flexibility and avoiding lock-in? Explain in two sentences.
> Draft three pointed questions I should email Vendor B before signing.
The first prompt did the reading. The follow-ups do the thinking *with* you.
If you have *one* document, the chat window is perfect. But what if you have 200 customer reviews and want each one labeled "positive," "negative," or "neutral," plus the main complaint?
Doing that by hand in chat is slow. This is where a short script helps. You do not need to be a programmer to read it:
from openai import OpenAI
client = OpenAI() # uses your API key
review = "The setup was confusing and support was slow to reply."
prompt = f"""
Classify this customer review as Positive, Negative, or Neutral.
Then give the main complaint in 5 words or less.
Review: {review}
"""
response = client.chat.completions.create(
model="gpt-5",
messages=[{"role": "user", "content": prompt}]
)
print(response.choices[0].message.content)In plain terms: this asks the AI the same question for each review automatically. Loop it over all 200 and you get a clean spreadsheet in minutes. The point is not the code. It is the mindset: *the same prompt that works once in chat can be repeated thousands of times by a script.*
Knowledge check
1. According to the lesson, what is the core skill being taught?
2. Why does the lesson insist that 'AI does the first pass, you do the final check'?
3. Why does asking the AI to cite the page or section of each point improve results?
4. Select ALL of the everyday tasks the lesson says large language models handle reliably.
Select all the correct answers.
5. Select ALL elements that make a summarization prompt 'strong' rather than 'weak' according to the lesson.
Select all the correct answers.
The difference between a frustrating result and a great one is usually structure, not the tool. Here is a reliable way to set up any analysis task.
Tell the AI who it is helping and why. "You are helping me brief a CFO who cares about cost and risk." This shapes tone and focus.
Upload or paste the actual documents. Do not describe them and hope the AI guesses. Give it the real thing.
Bullets? A table with named columns? A one-line summary? Say so. Ambiguous requests get ambiguous answers.
Add the line "If something is not stated, write 'not specified.'" This single instruction prevents most made-up details.
Spot-check key numbers and claims against the source. AI is your fast first reader, not your final authority.
Treat AI like a sharp intern who reads fast but sometimes overstates confidence. You would never forward an intern's first draft to your boss unread. Same here. Review, then send.
These actions are compiled in the role's Playbook.