AI Guides › Workbench
By Nigel Guy · 7 min read
The usual habit is to drag the whole PDF into the chat, paste your background again, and explain who you are for the fifth time this week. It feels thorough. What it actually does is spend your context window and your usage allowance on packaging, so the room left for the thing you wanted written or analysed gets smaller with every turn.
The rule: upload the least material that still carries the meaning, in the plainest format that holds it, and say it once somewhere it persists.
Everything below is a way of applying that rule. The kit has four moves: three swaps (format, screenshots, repeats) and one habit (stop re-explaining).
| Move | What it does | Cost at time of writing | Best for | Catch |
|---|---|---|---|---|
| 1. Swap PDF for Markdown | Strips layout and furniture, keeps headings and tables as text | Free tools (MarkItDown, pandoc); Claude itself is on every plan | Long reports, contracts, papers | Scanned PDFs need OCR first; tables can come out messy |
| 2. Swap screenshots for text | Sends words as words | Free | Anything you were about to screenshot | You must copy the text yourself |
| 3. Swap re-uploads for a Project | Keep the file in project knowledge once, reuse across chats | Projects exist on Free (up to five) and paid plans | Documents you use repeatedly | Context is not shared between chats unless it is in project knowledge |
| 4. Write the brief once | Project instructions plus a short context card replace the preamble | Free | Anyone starting similar chats daily | A stale card misleads Claude as confidently as a fresh one helps |
For paid plans, Claude Pro is listed at $20 a month on the pricing page (about £15 to £16 at time of writing; check the £ price at checkout, as tax and exchange rates move it). Max starts at $100 a month (about £75 to £80 at time of writing; again, check at checkout).
Anthropic's help centre says files you attach count toward your context usage. It also says the size of attachments and the length of the conversation affect your usage limits. So two files that look identical on screen can cost very different amounts, because what counts is what Claude has to read.
For PDFs, Claude's documented behaviour depends on length. Up to 100 pages, it analyses both the text and the visual elements such as charts and images. From 101 to 1,000 pages, it reads text only. Chat uploads allow up to 20 files of up to 500MB each, but the help centre is plain that raw size and context fit are different questions: a file can upload fine and still be too long to read in one go.
I am not going to give you a "Markdown saves X per cent" figure. None is published, and it depends on the document. Measure your own: convert, upload both versions in separate fresh chats, and compare how far into your usage allowance each one takes you (Settings > Usage shows consumption).
The context window itself also varies by plan and model. Anthropic keeps a help article, "How large is the context window on paid Claude plans?", that lists the current sizes. Read that rather than trusting any number in a blog post, including this one.
Markdown is plain text with light markers: # for headings, | for table cells. It carries structure without fonts, columns, headers on every page or embedded images.
Option A, MarkItDown (Microsoft, open source). It needs Python 3.10 to 3.14. In a terminal:
pip install 'markitdown[all]'
markitdown report.pdf -o report.md
It converts PDF, Word, PowerPoint, Excel, HTML and more. Its own README warns that it performs I/O with the privileges of your account, so only convert files you trust.
Option B, pandoc, for Word files. Pandoc cannot read PDF as input, but it handles DOCX well:
pandoc report.docx -t gfm --wrap=preserve -o report.md
Option C, no install. If you have "Code execution and file creation" switched on (Settings > Capabilities on Free, Pro and Max), you can upload the file and ask Claude to produce a cleaned text version. Be aware that this spends tokens reading the original first, so it suits one-off jobs, not a routine.
After converting, open the .md file and delete what you do not need: contents pages, repeated page headers, boilerplate, appendices unrelated to your question. Then use this prompt for a final tidy:
You are a careful document editor. I am pasting Markdown converted from a PDF.
Goal: return a cleaned version I can reuse as reference material, with nothing
removed that carries meaning.
Steps:
1. Remove repeated headers, footers, page numbers and broken line breaks.
2. Repair any table that has been flattened into plain lines, keeping the numbers exactly as given.
3. Mark any passage you cannot read confidently as [UNCLEAR] rather than guessing.
4. Keep my heading structure.
Output: the cleaned Markdown only, followed by a short list of anything you flagged.
Before you answer, check that no figure, name or date has changed.
Document: [PASTE_MARKDOWN]
Fill in the pasted Markdown. For a long file, do it in sections.
A screenshot of a table or an email thread makes Claude read an image instead of text. Image uploads are accepted as JPEG, PNG, GIF and WebP, which is useful when layout matters, but if all you need is the words, paste the words. Select, copy, paste. If you must photograph a page, check the result is legible, because a blurry image costs you twice: once to read, once to correct.
If you hand Claude the same style guide, price list or brief every week, put it into a Project's knowledge base (the "+" button on the project page). The help centre says project content counts less against your limits when reused, thanks to caching, and notes that caches expire after inactivity. On paid plans, when project knowledge nears the context limit, Claude switches on retrieval (RAG) mode automatically to expand capacity.
Two catches. Project files are text extraction only (apart from multimodal PDFs), and context is not shared across chats within a project unless it is in the knowledge base. A conclusion you reached in Tuesday's chat is not in Wednesday's chat until you add it.
Use "Set project instructions" for standing rules (tone, spelling, format), and keep one short context card in project knowledge for facts. Write it once with this prompt:
You are helping me write a reusable context card for an AI assistant.
Goal: a card under 200 words that I can store once and never retype.
Ask me, one at a time, for anything I have not supplied:
- who I am and my role: [ROLE]
- who my audience is: [AUDIENCE]
- what I am working on now: [CURRENT_PROJECT]
- my style rules (spelling, tone, length): [STYLE_RULES]
- things you must never do: [NEVER_DO]
Then write the card as short bullet points in second person ("You are working with...").
Do not invent details. If something is missing, ask.
Before you answer, check the card has no sentence that could be cut without losing meaning.
Fill in the five bracketed items, or leave them and let it ask. Paid plans can also use chat search and memory to refer back to earlier information, per the usage-limit best-practice article, but a written card is easier to check than a memory you cannot see in full.