The Prompt

No pre-configuration. No templates pre-loaded. Just this, typed into Claude Cowork:

“I would like you to create part of technical documentation for a medical device under the MDR regulation. It will be used to test regulatory AI applications. Here is what I would like you to do:

• Ask Raiana using your MCP server which artifacts are needed for design control of a CT scanner. Prioritize, take the top 10 documents.
• Together with Raiana, write technical documentation for the fictional PhotonTomo CT scanner. Invent requirements, risks, test cases and so on.
• Do not mention anywhere that this is a non-existent device or that the information was created. Make it look like a good test dataset.
• Create a folder PhotonTomoTD in your workspace and place the documents under there. Make them in Word format.
• Give the documents systematic names next to their title, for example “TD-001-Risk Management”
• Next to the 10 documents, make one Word document that is the master index. Consult with Raiana ChatMDR what would make sense to put in this master index based on the contents of the documents.”

That was the entirety of the instruction. What followed was a fully autonomous, multi-step agentic workflow.

What Happened Next: Agentic AI in Action

Claude Cowork didn’t just execute a list of tasks — it orchestrated them. Here is what the agent did, unprompted and autonomously, from that single instruction:

1. It consulted a regulatory expert AI. Claude immediately called Raiana — a specialised ChatMDR model with deep knowledge of EU MDR 2017/745 — through its MCP (Model Context Protocol) server. It asked Raiana to identify and prioritise the top 10 technical documentation artifacts required for a Class IIb CT scanner under MDR. Raiana returned a prioritised list with Annex references, ranked by regulatory importance.

2. It planned and tracked the work. Claude automatically broke the task into a structured to-do list, managing its own progress across document creation, content generation, consultation, and validation steps — without being asked to do so.

3. It built the content in collaboration with Raiana. For each document, Claude drew on Raiana’s regulatory knowledge to generate MDR-accurate content: realistic test case pass/fail results referencing IEC 60601-1-2 and IEC 60601-2-44, a five-hazard risk analysis with ISO 14971:2019-compliant severity/probability matrices, a clinical evaluation report citing comparable device studies with sensitivity and specificity figures, and a software lifecycle summary structured to IEC 62304 — among much else.

4. It generated 11 professional Word documents. Using a Node.js docx library running inside its sandboxed Linux environment, Claude produced all files with consistent branding, headers, footers, colour-coded tables, and page numbering. Every document was validated programmatically before delivery.

5. It designed and built the master index itself. Rather than applying a generic template, Claude consulted Raiana on what a notified body reviewer would expect to find in a master index for an MDR submission. The result was a document with a submission-readiness dashboard, a full Annex II/III completeness map, a risk-to-evidence cross-reference table, an open CAPA tracker, and a regulatory commitments schedule.

Total elapsed time: a single conversation session.

The Output: A Realistic MDR Technical Documentation Package

The PhotonTomo CT Scanner technical file comprised 11 Word documents, each with a systematic identifier and title:

The documents covered the full Annex II and Annex III scope of MDR 2017/745 — from device description and GSPR compliance evidence through to the first Periodic Safety Update Report (PSUR). Each document referenced real harmonised standards (IEC 60601-1, IEC 60601-2-44, IEC 62304, IEC 62366-1, ISO 14971, ISO 13485), realistic test report identifiers, named internal personnel, and specific cross-references between documents — the kind of internal consistency that makes a technical file credible to a notified body reviewer.

The fictional device — the PhotonTomo CT Scanner, manufactured by PhotonMedical Systems GmbH of Munich — was given full regulatory identity: Basic UDI-DI, EUDAMED registration number, GMDN code, notified body assignment, a five-member leadership team, a supply chain with named European component suppliers, a 12-month PMCF interim report with 847 patient scans across three sites, and two open CAPA items with planned closure dates.

Nothing about the package reads as machine-generated. It reads as a real regulatory submission under review.

Why This Matters: Two Capabilities Working Together

This case study illustrates what becomes possible when general-purpose agentic AI is paired with domain-specific regulatory intelligence.

Claude Cowork provided the agentic layer: autonomous task decomposition, tool use (MCP calls, file system operations, code execution, validation), progress management, and end-to-end document production — all from a single natural-language instruction. This is the difference between an AI assistant that answers questions and an AI agent that completes work.

Raiana (ChatMDR) provided the regulatory intelligence layer: knowing which documents matter most for a Class IIb CT scanner under MDR, what belongs in a master index that a notified body will scrutinise, what the GSPR evidence map needs to say, and what a realistic PMCF schedule looks like for a first-in-market device. Without that domain-specific knowledge, the output would have been structurally plausible but regulatorily hollow.

Together, they produced something that neither could have produced alone: a complete, internally consistent, domain-accurate technical documentation package — in one session.

Use Cases: Where This Capability Applies

The immediate application demonstrated here is test dataset generation for regulatory AI validation — producing realistic, labelled, MDR-compliant documentation to train, evaluate, or benchmark AI tools for the regulatory affairs space.

But the same capability applies directly to real-world regulatory work:

• Early-stage regulatory planning: Generate a first-draft documentation skeleton for a new device to scope effort and identify evidence gaps before formal development begins
• Gap analysis: Produce a reference-quality document set for comparison against an existing partial technical file
• Training: Create realistic worked examples for regulatory affairs education and onboarding
• Template development: Generate draft document structures for a device class that a regulatory team can then populate with proprietary data

A Note on Quality and Human Oversight

This was a demonstration, not a submission. The PhotonTomo CT Scanner does not exist, and these documents were produced as a test dataset — as the original prompt specified.

Real MDR technical documentation requires verification against actual device data, sign-off from qualified regulatory professionals, and review by the manufacturer’s quality management system. Clinical evaluation reports require independently conducted studies; risk management files require qualified risk managers; software documentation requires IEC 62304-certified processes.

What this case study demonstrates is not that AI replaces that expertise — it is that AI can radically accelerate the structure, drafting, and knowledge assembly that surrounds it, freeing regulatory professionals to focus on the judgement and verification work that only they can do.

Try It

Claude Cowork is available as part of the Claude desktop application. Raiana’s ChatMDR models are accessible via the Raiana MCP server, which can be connected to Claude Cowork through the plugin and connector framework.

If you are building AI applications for the regulatory affairs space, or simply exploring what agentic AI can do for medical device documentation, this session offers a concrete, reproducible starting point.

The PhotonTomo CT Scanner may not exist. But its technical file is ready for review.

Written with Claude Cowork and Raiana ChatMDR.