By the end of this hour you will have a Claude account, one Project holding your operation’s details, and one regulator document read back to you with the parts that touch your licence pulled out. You will also know the one document limit that decides whether Claude sees the tables in your PDF or ignores them completely.
This is for anyone in iGaming who reads regulator output as part of the job. Compliance, legal, country managers, product owners who get sent a consultation and have to say what it means by Friday.
The setup
You need a Claude account at claude.ai. Projects are available on every plan, including free, though free accounts are capped at 5 projects. Paid plans matter for one specific reason covered in step 3.
You also need one published regulator document. Use something already public: a UKGC consultation, an MGA circular, an ADM technical rule. Something like the UKGC’s warning to operators on identity checks is exactly the kind of output this workflow is built for. Do not start with an internal document, and do not start with anything containing player data.
Step 1: create the Project
Go to claude.ai/projects and click New Project in the upper right. Give it a name and a description.
Name it after the operation rather than the task. “Malta B2C compliance” works. “PDF summaries” does not, because you will end up putting unrelated work in it and the context stops being useful.
On Team and Enterprise plans you choose visibility at this point, private or shared with the organisation, with view and edit permissions.
Step 2: set the project instructions
Click Set project instructions. These apply to every chat inside the Project.
Keep them about behaviour rather than facts. Facts belong in project knowledge, covered in the next pillar in this series. Behaviour looks like this:
You are supporting a compliance team at a licensed online gambling operator.
Rules:
- Quote the exact article or paragraph number for every obligation you state, and quote the wording alongside it.
- If you cannot find the number in the document provided, say so. Never estimate one.
- Distinguish between what the document requires and what it proposes for consultation.
- Use GGR and NGR only where the source uses them. Never substitute one for the other.
- Flag every deadline with its date.Click Save instructions.
The instruction about article numbers is the one that earns its place. It gives you something checkable in every answer rather than a paraphrase you have to trust.
Step 3: add the document, with the page count in mind
The knowledge base sits on the right of the Project page. Click the + button to add content.
Before you upload, check how many pages your PDF has, because Anthropic’s documentation draws a hard line at 100.
Claude analyses both text and visual elements, meaning images, charts and graphics, in PDFs of 100 pages or fewer. For PDFs from 101 to 1,000 pages, Claude processes text only and does not analyse visual elements. Above 1,000 pages it will not take the file at all.
Regulator documents put a lot of load in tables, and plenty of those tables arrive as images rather than as text. A 140-page consultation with its fee schedule in a scanned table is a document where the fee schedule is invisible to the model.
The fix is to split the file. Cut the annexes and the tables into their own PDF under 100 pages and upload that separately.
Project files are capped at 30MB each. Files attached to a single chat rather than the Project can be up to 500MB, with a limit of 20 per chat.
Step 4: ask the question that produces a checkable answer
Open a chat inside the Project and ask for structure before content:
List every section of this document that creates an obligation for a
remote operator, with the article or paragraph number and the page.
One line each. No summary yet.You get a map. Now you can see whether the model found the annex, whether it found the fee table, and whether the article numbers run in a sensible sequence.
Then go deeper on the parts that matter to you:
For sections 4, 7 and 12 only: quote the obligation in full, state the
deadline, and state what a licensee has to change to comply.Step 5: check three references against the source
Open the original PDF. Pick three article numbers from the answer, one from the beginning, one from the middle, one from the last third. Find them in the source.
Do this every time, on every document, however good the previous answer was. It takes 2 minutes and it is the difference between a tool you can put in front of your DPO and one you cannot.
Where it breaks
Three documented limits will catch you out in the first week.
Visual content disappears above 100 pages
Covered in step 3, and it is the one that costs the most, because nothing in the interface tells you it happened. The model does not say “I could not see your tables”. It answers using the text it has. A consultation whose real content sits in a scanned annex will produce a confident, fluent, incomplete brief.
Context does not travel between chats
Anthropic states it directly: context is not shared across chats within a project unless the information is added into the project knowledge base. Work out a definition in chat one, open chat two, and it is gone. Anything you want to keep goes into project knowledge or into the instructions.
Non-PDF files lose their images
DOCX, XLSX, CSV, TXT and the rest go through text extraction only. Embedded images are not processed. A Word document where the licence matrix is a pasted screenshot uploads cleanly and arrives without the matrix.
The limits
This workflow reduces obvious errors in how you read a regulator document. It does not replace compliance sign-off, and no output from it should reach a regulator, a bank or a board paper without a named person having read the source.
Keep player data out of it entirely at this stage. Every workflow in this series that touches player-level data carries an anonymisation step written into the steps, and this one does not have one because it does not need one.
On the EU AI Act, an operator using Claude is a deployer rather than a provider, and the duties differ. Article 50 transparency obligations have applied since 2 August 2026, with a limited grace period to 2 December 2026 for marking obligations on systems already on the market. The deployer duty on AI-generated text covers text published to inform the public on matters of public interest, and text that has had genuine human review and editorial control is exempt. Internal compliance briefs are a different question from published copy, and your legal team owns where that line sits for your business.
AI for iGaming is a recurring series from The iGaming Europe. Bartosz, our Head of Content, shares the prompts, setups and workflows he has tested on real iGaming work. One workflow per issue, including what went wrong. No tools to buy, no affiliate links.
Source: Anthropic Claude Help Center, European Commission
