Research, writing and analysis with AI

In part 1 you learned how the model works, how to give it context and how to write instructions. In module 6 you organized that into projects and gave it access to your email and files. This module is a set of recipes for the work of an ordinary Tuesday, each with three parts: when to use it, how to ask for it, and what to check before you use the result.

There are five: research with sources, writing, data analysis, documents and meetings, and building a routine. In 4D terms, this module is Delegation in practice: deciding which task you hand over, in what form, and where your review sits.

How do I research with AI without it inventing the sources?

A model without tools answers from its training memory, with the knowledge cutoff and the tendency to make things up you saw in module 1. With web search, it checks the internet before answering and cites what it found. With deep research (Claude calls it Research), it plans, runs many searches and delivers a report with citations. The practical question is which one to use.

• Web search for direct questions that one or two lookups settle: what a regulation says, who runs a company, what a service costs today. As of October 2026, depending on your version of the Claude app you either turn it on from the "+" menu or Claude searches on its own when the question calls for it; to force it, write "search the web" in your prompt.
• Deep research for broad questions: comparing three suppliers, understanding how a regulation changed, mapping a new market. Anthropic recommends it when the task needs five or more lookups; it takes one to three minutes, is on paid plans, requires web search to be on, and uses up your usage limits faster than a normal conversation.
• Neither when the answer is in your own documents. For that you do not search the internet: you upload the file or use the project from module 6.

The recipe, in either mode: always ask for sources, and ask for the date of each one. A fact without a source is the model’s opinion; a fact with a three-year-old source may be out of date. Then the step nobody wants to do but that separates serious work from risky work: open every source and confirm it says what the text claims. Claude’s own documentation says the same: the citations are there for you to verify.

Research [TOPIC] so that I can [WHAT I NEED IT FOR: for example, decide whether to open a branch in Barranquilla, reply to a customer, update our cancellation policy].

Context: [TWO OR THREE LINES ABOUT YOUR BUSINESS AND WHAT YOU ALREADY KNOW ABOUT THE TOPIC].

Rules:
1. Use web search. Prefer official sources (public bodies, the manufacturer’s own site, published regulations) over blogs and news.
2. Every concrete claim (figure, date, regulation, price, name) carries its source. If you cannot find one, say so instead of filling the gap.
3. If two sources contradict each other, show me both and tell me which is more recent.

Deliver:
- A summary of at most 10 lines answering my question.
- A claims table with columns: claim, source (link), date of the source, confidence (high, medium, low) and why.
- A list of what you could not verify.

How do I write emails, proposals and policies without sounding like a robot?

Writing is where most people use AI and where it shows most when they use it badly: correct, long, generic text in a tone that is not yours. The difference between a useful draft and one you have to rewrite is almost never the model; it is what you gave it. Everything in this section starts from two things you already have: the context document from module 2 and the prompt structure from module 4.

• Emails. Give it the previous thread (anonymized if it contains other people’s data), what you want the reply to achieve and what you will not concede. Set a word limit. Check that it promises nothing that is not in the context: dates, discounts, deadlines.
• Proposals and quotes. Paste an earlier proposal you liked as the example of structure and tone, the context document and this client’s details. Ask it to put everything that needs your confirmation (prices, timelines) in square brackets. Check every bracket before sending.
• Policies and procedures. Describe how the process works today, with the real exceptions, and ask for a draft with numbered steps, owners and what happens when something fails. Review it with the person who runs the process: they spot what does not fit in a minute.
• Reports. Hand over data you have already verified and the conclusion you reached; ask it to write, not to conclude. If you do want it to conclude, ask it to separate "what the data shows" from "what I interpret".

Two habits change the result more than any template. First: edit your own text instead of generating from scratch. Write the email yourself in five clumsy lines, then ask it to "improve clarity and tone without adding ideas or changing how I speak". What comes out is yours, polished. Second: give it examples of your voice. Two or three pieces of your own writing that you like, pasted as examples, teach more than any description of "friendly but professional". Those examples live in the prompt library from module 4.

Can I upload a spreadsheet and ask for an analysis?

Yes, and it is one of the uses with the best return, with one condition that comes from module 1: a model that "adds in its head" produces a plausible number, not a correct one. The fix is to have the assistant write and run code to calculate instead of estimating. As of October 2026, Claude calls this code execution and file creation; it is available on every plan, including the free one, and you turn it on under Settings, Capabilities. It is also what you need to upload Excel files (.xlsx); a CSV uploads without it, but the calculation should still be done with code.

• When. A sales report by branch, the month’s booking list, a project’s expenses, an export from your accounting system. Any table you would ask questions of if you had a free afternoon.
• How. Before uploading, remove the red columns from the traffic light in module 5: names with ID numbers, phone numbers, bank accounts. The analysis almost never needs them. Then say which decision the analysis serves, not just "analyze this". Ask it to use code for the calculations and to show what it calculated and how.
• What to check. The grand total against a total you already know (from your system, or a quick sum in Excel). The number of rows it processed against the number of rows in the file. And one figure you know by heart (sales in a given month, spend with a given supplier). If those three match, the rest probably does too; if one fails, nothing in the report gets used until you understand why.

The same tool works in the other direction: the assistant can create the Excel file with the clean table and formulas, the Word document with the report or the PDF to send. According to Claude’s help center, the limit is 30 MB per file and these tasks use more of your usage limit than a normal chat. For data that arrives as PDFs or photos (invoices, delivery notes), the principle is the same, and we cover it in this guide to data extraction.

Attached is [FILE NAME], which contains [WHAT IT IS: for example, sales by product and branch for the last 12 months]. The columns are: [LIST OF COLUMNS AND WHAT EACH ONE MEANS].

The decision I want to make with this analysis is: [DECISION: for example, whether to close branch B or change its opening hours].

Rules:
1. Use code for every calculation; do not estimate any figure.
2. Before answering, tell me how many rows you read, the date range and the grand total of [MAIN COLUMN], so I can compare them with my system.
3. If there are empty, duplicate or oddly formatted rows, list them and tell me what you did with them.
4. Separate "what the data shows" from "what I interpret" and flag any assumption you made.

Give me: the three findings most relevant to my decision, each with the figure that supports it and how you calculated it, and a summary table I can paste into a report.

How do I get value out of long documents and meetings?

Reading is slow and the model reads fast. That is the advantage, and also the risk: a summary that sounds complete may have skipped the clause that mattered. Four recipes, each with its check:

• Summarizing a long document. Upload the PDF (Claude reads up to 1,000 pages of text, and charts and tables in documents of up to 100 pages), say who the summary is for and which decision it feeds, and ask it to cite the page or the exact passage for every point. Check: open three citations at random and confirm they say that.
• Comparing two versions of a contract. Upload both and ask for a table: clause, what it said before, what it says now, who the change favors. Check: read for yourself the clauses marked as changes in your favor; a model inclined to please can read them optimistically. This does not replace your lawyer; it prepares you for the meeting.
• Extracting obligations and dates. From a contract, a tender or a set of regulations, a table with obligation, who fulfills it, deadline, consequence of missing it and the clause where it lives. Check: open two or three clauses.
• From transcript to minutes. Paste the transcript (tell participants the meeting is being transcribed) and ask for decisions taken, tasks with owner and date, open topics and questions nobody answered. Check: send the draft to the attendees before calling it the minutes; anyone who was in the room spots what the model misread in a minute.

In every case, the Discernment question is the same: did the model work only with what I gave it? One instruction helps a lot: "use only the attached document; if something is not in it, say so". And the traffic light still applies: a contract or a transcript with third parties’ personal data is yellow or red material depending on your plan and on what it contains.

Attached are two versions of [TYPE OF DOCUMENT]: version A is [THE ONE I SENT / THE ORIGINAL] and version B is [THE ONE THEY RETURNED / THE NEW ONE]. I am [YOUR ROLE: for example, the landlord, the supplier, the customer].

Work only with the text of the two documents. Assume nothing that is not there.

Give me:
1. A table with every difference: clause or section, text in A, text in B, what changes in practice, who it favors (me, the other party, neutral).
2. A list of everything that was in A and is gone in B, even a single word.
3. The three differences I should most review with a lawyer or an accountant, and why.
4. Anything that did not change but I should still look at, if you see it.

This is the transcript of a meeting about [TOPIC] with [PARTICIPANTS OR ROLES]. Turn it into minutes with these sections:

1. Decisions taken (one line each, in the past tense: "approved", "dropped").
2. Tasks: what, owner, date. If no date was said, write "no date" instead of inventing one.
3. Topics left open.
4. Questions someone asked that nobody answered.

Use only what is in the transcript. If something is ambiguous, mark it [CHECK] instead of interpreting.

Transcript:
[PASTE THE TRANSCRIPT HERE]

Which three workflows should I start with this week?

Five recipes at once is zero recipes. Experience with small teams says a workflow becomes a habit when it is repeated three or four times in one week with the same saved prompt. Pick three, in this order if you do not know where to start:

1. 1Editing your own text. It is daily, low risk, and it teaches you quickly how the model writes and how you write. Emails and long WhatsApp messages are the natural place.
2. 2One sourced research task a week. Something you were going to look into anyway: a supplier, a regulation, a competitor. With the claims table and the links opened. It trains your eye for where it invents.
3. 3Meetings or documents, depending on your week. If you have meetings with transcripts, the minutes. If contracts or proposals come back with changes, the comparison. One of the two, not both yet.

Spreadsheet analysis comes in the second week, once checking is a habit. And one rule for the routine: every prompt you use twice goes into the library with the date it last worked well. When the model or the app changes (and they change every few months), you will know what to test again.

Everything in this module is something you do yourself, in a conversation, when you need it. The next step is for some of these tasks to happen without you asking: a news digest every Monday, data extraction when an invoice arrives. That is module 8, "Automations: when AI works on its own", and it starts with deciding what is worth automating and what is not.