Give Claude the mess

I used to type "summarize my last meeting" like Claude had been in the room. It would invent a tidy agenda. Action items. A person named Sarah who was not there. I nodded at it for a second because it sounded like a meeting. Then I remembered I had the notes in my phone the whole time.

If you did step 1 you already know the chat works. This step is the one people skip because the real material is ugly. The notes have half sentences. The email thread has three people talking past each other. The product page is a paragraph you hate. That is the useful stuff.

Asking Claude to "write a recap" with no source is how you get a recap of a meeting that never happened. Same for "make this landing page better" when this is not in the box.

Put the pile in the box

You have two honest ways to do this on claude.ai.

Paste. Copy the email thread, the notes, the Google doc text, the rant you wrote at 11pm. Drop it under a one-line ask.

Or attach the file. In the chat box there is a + button on the lower left. Official help says you pick Add files or photos, or you drag the file onto the window. You can paste an image from the clipboard too. I still forget the + is there and paste a wall of text. Both work.

Claude's upload help lists the usual office pile. PDF, Word (DOCX), CSV, TXT, HTML, a few other text formats, plus JPEG, PNG, GIF, WebP. Spreadsheets in XLSX need code execution and file creation turned on in Settings, under Capabilities. I would not start with a 40-tab workbook. Start with the email.

Limits, because people will try to attach a whole life. In a single chat you can put up to 20 files, 500MB each. PDFs go up to 1000 pages. After 100 pages Claude stops looking at the pictures and charts and just reads text. Project files are a different, smaller per-file cap, 30MB, and that is for later when you make a Project.

What the mess looks like when it is real

Last month I had a landlord thread. First message was "routine inspection." Second message added a date I would be at work. Third message, from a different person, mentioned a lockbox. I could have asked Claude to write a tenant rights essay. I pasted the three emails and wrote: what are they asking, what did they change, draft a reply that offers Saturday morning and asks them not to use a lockbox without me there.

The reply was usable after I cut two sentences. The useful part was not the polish. It was Claude lining up what each sender actually said, because I had stopped reading carefully after email two.

Meeting notes work the same way. Mine look like this.

jen: budget? wait for Q4 me: no, we said August sam: page still says coming soon ?? who owns the FAQ parking lot: refunds

That is a mess. It is also better context than "we had a product meeting." I paste that and ask for a list of decisions, a list of things we did not decide, and a checklist I can drop in Slack. I do not ask for minutes in full sentences. I was there. I need the shape.

Half drafts are the other pile I keep avoiding. I had a product page that started "Our platform empowers teams" and then a feature list I copied from a deck. I pasted the whole block and said tell me what a tired person still would not know after reading this, then rewrite the first 80 words like a human who has used the thing. I put my own sentences back in. Fine. The point is Claude reacting to your words, not a genre.

If you are embarrassed by the notes, good. That usually means they are specific.

Ask for one job, not a cleanup service

Once the source is in the chat, give Claude a single job.

Read this thread and tell me what they want by Friday.

Turn these notes into a checklist. Mark anything that is a guess.

This is a half draft. Keep my examples. Kill the slogans.

Do not stack "also make it a blog post and a script and a FAQ." You will get a blur, and you will not know which part came from your notes.

I like a two-beat prompt after the paste.

First: what is actually in here. Quote the lines you are using.

Second: now do the job. Draft, list, outline, whatever.

The quote part is how you catch the fake Sarah. If Claude cannot point at a line, it is filling. That habit matters more in step 4, but you can start it now.

If the file is a PDF of a policy or a lease, say which pages you care about. Anthropic even tells you to use the page numbers in the viewer, not the printed numbers on the page. I learned that after Claude and I argued about page 4.

The empty summarize request

Here is the waste version, because I still do it when I am rushing.

"Summarize this."

This is not in the chat. There is no attachment. Claude will still write a summary. It will be smooth. You will feel briefly organized. You have no way to know if it is true.

Same pattern with "write an email to my landlord about the inspection." No dates. No names. No thread. You get a template. Templates are how you agree to a lockbox you did not mean to agree to.

If you do not have the file in front of you, go get it. Take a photo of the paper and paste the image. Forward yourself the thread and copy it. Type the three bullets from memory and label them as memory, which at least keeps Claude from treating your guesses as quotes.

When a longer thing opens on the right

Sometimes you ask for a full rewrite or a one-pager and Claude puts it in a dedicated window next to the chat. That is an Artifact. Official docs say it shows up for substantial standalone content, documents, code, a simple page, a diagram. On my machine it is the first time the product feels like more than a chat log.

You can ask it to change a section. You do not need to publish anything.

Artifacts need Code execution and file creation on, under Settings then Capabilities. If the window never appears, that toggle is the first thing I would check, before you decide Claude "cannot write documents."

Still, do not go make an Artifact for fun tonight. Get the mess in, get one useful object out. If the useful object is just a tighter email in the main thread, that is a win.

What I do not give it

Passwords. Bank stuff. A customer's full file if I would not put that in a random web form. Use your own judgment. I am not your lawyer.

I also do not dump six months of Slack. The model has a context window. The free plan has a usage window that resets every five hours. A giant paste that you do not need will burn both. Pull the thread that matters.

If the work is really a pile of code and files in a folder, stop trying to recreate the repo in the chat. That is Cursor. If you need to know whether a public thing still works this month, Grok is the search-y one. Claude is the place I bring the private mess I already have.

Next up is the boring middle, which is outlines and rewrites and turning a rant into something you can send. You keep the decision. Claude does not.

If you only do this step once, paste the actual notes. Ask it to show the lines it used. Then ask for the checklist or the draft. The tidy summary with no source can wait forever.