Claude does way more than writing: 4 starter prompts

Worth knowing why this matters right now: AI literacy came in at number one on LinkedIn's Skills on the Rise 2025 list of the fastest-growing skills in the US, described as the ability to understand and use AI tools for business purposes (LinkedIn News). Being able to draft an email with AI is not that skill. Handing it real work is.

1. Turn a spreadsheet into an actual analysis

This is the one people are most surprised by. When you upload a CSV or an Excel file and ask a numbers question, Claude does not eyeball the rows and guess. It writes and runs code on your file inside the conversation, then reports what the code found, which is why it can handle a file far too long to read (Anthropic's analysis tool announcement).

1. 📁 Export the raw file. Not a screenshot, not a pasted table. Download the CSV or XLSX from wherever the data lives so the column headers and every row come with it.
2. 🧹 Fix the header row. One row of clear column names at the top, no merged cells above it, no blank rows. Ninety percent of confusing answers come from a messy header, not from the model.
3. 📎 Attach it and ask. Use the paperclip in a new chat, upload the file, then paste the prompt below. Uploads work on every plan, up to 30MB per file (upload files to Claude).
4. 🔍 Check its working. Pick one number from its answer and verify it in the spreadsheet yourself. Do this the first three times and you will know exactly how far to trust it after that.

Here is my [describe the data, for example "last 12 months of sales by product and region"]. Before you analyse anything, read the file and tell me what columns you actually have, how many rows, what date range it covers, and anything obviously broken like blanks, duplicates or values in the wrong format.

Then run the analysis: the top three trends over the period, the biggest outliers and what makes them outliers, and any relationship between columns that I probably have not noticed. Show me the numbers behind each claim, not a summary of them.

Finish with the three things you would look at next if this were your business, ranked by how much money or time is attached to each. If a pattern is based on too few rows to be meaningful, say so plainly instead of presenting it as a finding.

2. Map a decision instead of asking for advice

Asking "what should I do" gets you an opinion shaped by however you framed the question. The move is asking for the structure underneath the decision: what each option costs, what has to be true for it to work, and what would tell you early that it is not working. You keep the call, and you make it with the tradeoffs written down.

1. 🗒️ Brain-dump the mess. Paste everything: the constraints, the numbers, the politics, the deadline, the thing you are embarrassed about. Context is the whole input here, and a tidy summary throws away the parts that actually decide it.
2. 🎯 Name the real decision. One sentence, in the form "I need to decide whether to X or Y by [date]". Vague inputs produce advice-column answers.
3. 🧾 Ask for structure, not a verdict. Use the prompt below. It forces risks, assumptions and reversal costs into the open.
4. 🔁 Then argue with it. Reply with "argue the opposite case as strongly as you can" and see which of your reasons survive. That second pass is where most of the value is.

I have a decision to make and I want the structure of it, not your recommendation yet. Here is the full context, including the messy parts: [paste everything].

Lay it out for me as a table of the realistic options, and for each one give me: the cost in money and in time, what has to be true for it to work, the main risk, how reversible it is if I am wrong, and roughly when I would know either way.

Then tell me the single assumption the whole decision rests on, and the cheapest test I could run this week to check that assumption before committing. Separate what you know from my context from what you are inferring, and if there is an option I have not considered, add it to the table and say why it belongs there.

3. Build a clickable mockup, not a picture of one

This is the part that gets confused with image generation. Claude can build the real thing: a working page or small app that renders in a panel beside the chat, where you can click buttons, type in fields and see what actually happens. Those are artifacts, and you can publish one as a link that anyone can open without a Claude account (Anthropic's artifacts guide).

1. 🎯 Decide what you want to test. A landing page, an internal dashboard, a pricing calculator, an onboarding flow. Name the one thing a viewer should be able to do with it.
2. 🎨 Bring your brand. Paste your hex codes and font names in the first message. Two minutes here is the difference between a generic template and something that looks like yours.
3. 🖱️ Prompt, then click everything. The artifact opens beside the chat. Actually use it, find the dead ends, and say "the pricing toggle does nothing, make it switch the numbers" in plain English.
4. 🔗 Publish and send it. Use the publish button in the artifact panel to get a link you can put in front of a client or a teammate (publish and share artifacts). Feedback on a thing people can click beats feedback on a description every time.

Build me a working, clickable artifact for [describe the page or tool, for example "a landing page for a 4-week course for beginners"]. I want something I can actually interact with in the preview, not a static image or a description.

The audience is [who], the one action I want them to take is [the action], and the sections I need are [list them]. Use my brand: [hex codes] for colours, [font] for headings and [font] for body, and keep it clean rather than busy. Make it responsive so it holds up on a phone.

Every interactive element has to genuinely work: buttons change state, forms validate and show a confirmation, toggles switch what is displayed. When it is built, list the assumptions you made that I should confirm, and give me three specific things about it you would test with a real user before I commit to this direction.

4. Automate the boring repeat job

Anything you do the same way every week is a candidate: renaming and sorting files, merging four exports into one report, pulling the same numbers into the same template. You describe the task in plain English and Claude writes the thing that does it. You do not need to read the code, but you do need to test it on copies first.

1. 📝 Describe the task like a handover note. What the input looks like, what the output should look like, and every rule and exception. Write it as if a new starter has to do it without asking you a question.
2. 🧪 Give it a real sample. Upload two or three example files so it can see your actual formats instead of designing for an imaginary one.
3. 📁 Test on copies, always. Duplicate the folder or file first and run it there. This is the rule that keeps an automation experiment from becoming an incident.
4. ♻️ Turn the keeper into a skill. Once it works twice, ask Claude to package the process into a reusable skill. Install it in the app under Customize, then Skills, then Add, then Upload a skill. Skills are available on every plan, Free included (creating custom skills).

I do this task manually every [week / month] and I want to stop: [describe the task exactly as you would hand it to a new starter, including every rule and exception].

The input looks like this: [describe or attach a sample]. The output I need looks like this: [describe or attach an example of the finished thing]. Before you build anything, ask me the questions you need answered to get the edge cases right, and tell me what you would do with a file that does not match the expected format.

Then build it and hand it back in a form I can actually run without being technical. Tell me in plain English exactly what to click or type, step by step, from a completely fresh start. Include a way to test it on copies of my files first, and tell me what a successful run should look like so I know it worked.

What separates the two groups

The people getting real leverage out of Claude are not writing cleverer prompts. They are being more specific: naming the file, the date range, the constraint, the audience, the format they want back. Every prompt above is long for exactly that reason. Vague in, generic out, every time.

Proof

LinkedIn Skills on the Rise 2025, AI literacy at number one
Anthropic, introducing the analysis tool
Create and edit files with Claude
Upload files to Claude
Use artifacts to visualise and create AI apps
Creating custom skills