Moving a large digital transformation project into Claude Cowork — how would you structure the project knowledge?
Hi everyone, I’ve decided to move the management of a large digital transformation project into Claude Cowork, and I’ve hit a problem much earlier than expected: how should I structure the project knowledge so it can scale?
For context, this is a fairly complex project:
- ~2 years of accumulated project data
- Confluence + local files
- Jira, Slack and Microsoft 365/Teams
- ~15 vendors
- Teams distributed globally
- Decisions, risks, business requirements, actions, dependencies, etc.
- Information (status, scope) constantly changing through Slack conversations, Teams / Outlook messages, Jira updates and Confluence pages
At this scale, I’ve reached the point where I no longer have good visibility over the whole project.
So: Last week I installed Claude and got the necessary admin access to Microsoft 365, Jira, Confluence and Slack. Initially I was excited about connecting everything, but I’ve quickly realised that giving an AI access to the information isn’t the same as giving it a good information architecture.
That’s where I’m stuck. How should the local project structure work? I’m seeing conflicting advice about the ideal architecture, particularly around Markdown. For example, should I maintain separate files such as:
/project PROJECT.md STATUS.md DECISIONS.md RISKS.md REQUIREMENTS.md DEPENDENCIES.md ACTIONS.md
Or should these be broken down further into folders and smaller files as the project grows?
I’m particularly interested in file size and boundaries. If DECISIONS.md, for example, eventually contains hundreds of decisions accumulated over several years, is that still the right approach? Or should each decision become its own Markdown file, perhaps with an index/summary file on top?
The same question applies to risks, requirements, actions, vendor information, workstreams, etc.
I’m less concerned about automation at this stage. Anthropic’s documentation around Skills gives me a reasonable starting point for that.
What I want to get right before automating anything is the underlying project architecture.
For people using Claude/Cowork on genuinely large, long-running projects:
- How do you structure your Markdown files and folders?
- How large do you allow individual .md files to become before splitting them?
- What information do you keep in Markdown versus leaving in Jira/Confluence/Slack as the source of truth?
- And how do you prevent the local project knowledge base from becoming another giant information dump that the AI has to rediscover every time?
Thank you, Clement