To work productively with Claude, you cannot avoid the invisible half: configuration. Which instructions stand in the profile. What lies in the projects. Which skills are active. Which connectors are connected. Which folders on my computer are shared. These settings determine how competent Claude appears every day — and they are underexposed in most guides. This article closes that gap from my own practice.
Instructions: the foundation
The most important one-time setting in Claude.ai and Claude Cowork is the personal instruction. It lies in the profile area and is carried along in every new conversation. If you leave it empty, you get Claude in its default state: competent, polite, contextless. If you fill it in, you get Claude as a mentor who knows your professional profile, respects your values, and matches your style.
My instruction spans thirty to forty lines and contains five blocks. First block: who I am professionally and privately. Second block: which roles I play (for me, that’s three — Application Manager at SPL Tele, founder of Lumi-Systems.io, psychology student). Third block: which core competencies I have and in which languages I communicate. Fourth block: which values are non-negotiable — for me, life, freedom, learning. Fifth block: how Claude should communicate — for me, academically precise, in simple language, without phrases like “does that make sense to you”.
These five blocks take twenty minutes to set up and save countless repetitions over the following months. If you have to re-explain your background in every new conversation, you don’t have a profile instruction yet. If you wonder why Claude sometimes answers in English in one session and in German in another — same story.
Projects: the bridge between sessions
Claude.ai and Claude Cowork know the concept of projects. A project is a persistent workspace with its own files, its own instructions, its own conversation history. Once you’ve set up a project, you can reopen it anytime, and Claude knows what we were working on last.
I have projects for every major work area. Aktera Connect has a project with the current architecture specification, data model, and open decisions. Lumi AI News has its own project where the source-of-truth file for Feedzy imports lives and the tag-filter concept is documented. My psychology studies have a project with the APA style guide and current module learning objectives.
The effect is enormous. When I go into the Lumi AI News project and write “let’s write a new editorial on the Verena Becker line”, Claude knows in seconds what the Verena Becker line is, which editorials we’ve already published, what tone fits. Without projects, every session starts at zero. With projects, every session starts exactly where the last one left off.
Uploading files: the underestimated lever
When you give Claude a PDF, an Excel spreadsheet, an image, a code file, you shift the discussion to a different level immediately. I can upload Verena Becker’s PDF presentation and say “read that, tell me the five most important points, and compare with what I already know”. I can upload an Excel with our SPL Tele application data and say “show me which applications could be redundant”. I can upload a screenshot of an accounting report and say “extract the numbers, make me a clean table”.
Three practical recommendations for file uploads. First: prefer uploading PDFs directly rather than copying — Claude understands the structure (headings, lists, tables) better than ASCII conversion. Second: screenshots are often the fastest bridge — if a tool or browser tab has no API, a screenshot is the acceptable input. Third: don’t upload credentials in files — they become part of the conversation context, and conversation contexts are not made for secrets.
Skills: the growing library of specialized competencies
Skills are a concept that has gained importance with Claude Cowork and Claude Code. A skill is a bounded competency package — instructions, possibly helper scripts, often its own knowledge base — that Claude activates as needed. Anthropic provides a range of pre-built skills (Word documents, Excel tables, PowerPoint, PDF handling, proprietary market report skills), and every user can add their own.
My own skills have grown over months. A “Rupp Frontend Design” skill that creates frontend components according to my ALEM design principles. A “Rupp Security Audit” skill that automatically checks every file creation for OWASP Top 10 issues. A “Lumi XWiki Bridge” skill that generates PowerShell commands for maintaining our XWiki. A “Meeting Transcription Analysis” skill that turns voice transcripts into management-ready reports.
Skills have a hidden lever: they automatically activate when Claude recognizes that the topic matches. I don’t have to say “use the frontend design skill” — when I ask for a React component, the skill loads. This saves explanation effort and ensures quality standards remain consistent.
Connectors: connections to the real world
Connectors are what distinguishes Claude Cowork from Claude.ai. While Claude.ai is essentially a conversation tool that works with files and web access, Claude Cowork has access to a growing range of connections to external services. Microsoft 365 (Outlook, Calendar, Planner, OneDrive, SharePoint). Google Workspace. Slack. Notion. Anytype. WordPress. Custom MCP servers.
I have active: Microsoft 365 (for the SPL Tele context, with strict separation), Google Workspace, a custom MetaMCP server for the Lumi Systems infrastructure, a WordPress connection for ainews.lumi-systems.io and lumi-systems.io itself, multiple custom connectors for Coolify and Hetzner. Each of these connections is a verified channel: I decide which app gets which permission, and I can revoke connections anytime.
One thing beginners often overlook: connectors are not just “read”. They are also “write”. If Claude Cowork is allowed to see my Planner, she can create tasks. If she’s allowed to connect Outlook, she can draft emails. This write permission is power, and power needs clear boundaries. My principle: every new connector permission gets a mental “worst-case” check. What could go maximally wrong? If the answer is “nothing dramatic”, the connection goes through. If the answer is “data leak or email sent in the wrong tone”, I think again.
Local folder sharing: the interface to the computer
Cowork runs on your own desktop and thus has the ability to read and write local folders — if I allow it. This permission is granted per folder, visible, and revocable. For me, shared are: a work outputs folder where all Cowork results land; a Lumi Systems directory with current configuration files; a temporary exchange folder for file operations.
The effect is immediately noticeable. When Claude Cowork generates a new WordPress plugin, the ZIP lands in the outputs folder — I click on it and install it in a minute. When she produces an Excel evaluation, I can open it immediately in Microsoft Excel. Conversely, if I put a PDF in the exchange folder, I can ask Claude “check out the new file” — no upload, no extra click.
One discipl