Claude
at work.
Three sessions in. You can chat, you can connect tools, you can write a Skill. Today we operationalise it. Projects, shared playbooks, Office 365, and how to do all of it without leaking the things we cannot afford to leak.
From using Claude to running on Claude.
Each session has moved one notch deeper. The shift today is from individual hacks to a team operating model. Pin context once, share playbooks, route the work through Claude instead of around it.
Ask better questions
Prompting basics. One off conversations.
Files in, content out
Drag PDFs, get drafts. Local memory.
MCP and Skills exist
Wire SharePoint, write a playbook.
Team rhythm on top
Projects, shared Skills, secure defaults.
Goal of today. Leave with one Project set up, one Skill shared, one Connector enabled.
Projects are persistent workspaces.
A Project is a folder Claude remembers. Drop in the docs that matter, write the instructions once, and every chat inside the Project starts already briefed. This is the single highest leverage move on Claude Pro and above.
The reference shelf
Upload up to 200K of tokens of context. Tone of voice docs, our PIAM datasheets, last three RFP responses, the customer pack template. Claude sees them in every chat.
The standing brief
"You are a writer for ID-ware. Use UK English. Quote our datasheet language. Always include a risks section." Claude follows it without you re-typing.
Anything you do twice
Bids. Customer support packs. Board updates. Monthly reports. The second time you copy paste the same context, that is a Project.
Enterprise rules apply
On Team and Enterprise, Project content is not used to train Anthropic. Stays inside our workspace. Owners control sharing and retention.
Skills you actually build.
A Skill is a small folder that says "when X comes up, here is how we do it". Claude loads it on demand. You do not need to be an engineer to write one. Describe the job in plain English and Claude scaffolds the rest.
SKILL.md + extras
A description at the top tells Claude when to use it. The body is the method. Optional references, templates, scripts sit alongside. That is the whole structure.
Ask Claude to write it
"Turn this conversation into a reusable Skill called rfp-response. Save when we should use it, the steps, and the gotchas." Refine on next use.
Word, Excel, PowerPoint, PDF
Anthropic ships file-creation Skills. Ask for "a Word document" and you get a real .docx, not markdown. Same for .xlsx and .pptx with charts.
A note on slash commands. The Claude.ai chat does not have a slash command menu. The "@" key attaches a file, a Connector resource or a Skill. The model picker, web search and extended thinking are buttons. /commands are a Claude Code thing, a different product we are leaving aside today.
Skills are how we scale taste.
If you figure out a great way to do something once, the team should never figure it out again. Treat Skills like code. Source controlled, reviewed, versioned. They are the smallest shippable unit of "how we work".
Org-wide, admin-gated
On Team and Enterprise, an admin can publish a Skill so every account sees it. House style, RFP template, security pack format. Update once, everyone benefits.
Third party Skills are extensions
If a Skill comes from outside ID-ware, read it. What does it tell Claude to do? Which files does it touch? Pin a version. Do not auto-update.
100+ ID-ware Skills
Aikido to Jira. Cognito M2M setup. Customer onboard. PIAM threat model. Each one bottles a runbook so you can stop relearning them.
Onboarding compresses
A new joiner inherits the team's playbook on day one. Not a wiki to read. A folder Claude already follows when the right question lands.
Connectors that work today.
Anthropic ships first party connectors for the obvious ones. Microsoft 365 is a notable absence. Here is the honest map and what to do about it.
Click and connect
Google Drive, Gmail, Calendar, GitHub, Linear, Asana, Jira, Confluence, Notion, Box, Salesforce, HubSpot, Zapier. Admin enables, user OAuths.
No first party yet
No native SharePoint, OneDrive, Outlook or Teams connector from Anthropic. Three routes work. MCP servers, Power Automate, upload by hand.
Bridge via Graph API
Microsoft and the community ship MCP servers for Graph, SharePoint, Outlook. Add via Claude Desktop, Settings, Developer. Start with read-only scopes.
Admin gate. On Enterprise, Connectors are an admin setting, not a free for all. Enable per workspace, scope to read where possible, log activity. Treat every new connector as a data egress route worth reviewing.
Three ways to put Claude on your M365 stack.
Drag and drop
Up to 30 MB per file, 20 files per message. Reads cells and formulas in .xlsx, speaker notes in .pptx, headings and tables in .docx. Quickest start.
Live, two way
Microsoft Graph MCP. Claude searches SharePoint, reads a workbook, drafts a reply in Outlook. Approve writes manually until you trust the loop.
Triggered flows
Cloud flow fires the Anthropic API when an email lands or a SharePoint item changes. Best for "always on" jobs like inbox triage or supplier intake.
- !What Claude sees in your files. Excel formulas as strings, not recalculated. PowerPoint speaker notes yes, animations no. Word comments and track changes sometimes lost. Accept changes before uploading if you want clean text.
- →Artifact vs paste-back. Ask for an artifact (real .xlsx or .pptx) when the file is the deliverable. Ask for pasteable markdown when you are editing inside the existing deck or workbook and want to keep the master.
Prompts that earn their keep.
Spreadsheets
Audit this workbook for circular refs, hard-coded numbers, inconsistent units.
Suggest an XLOOKUP that pulls Region from Sheet2 keyed on Customer ID.
Clean this CSV. Trim, ISO dates, dedupe on column A, return CSV.
Decks
Outline this 30 page doc as a 10 slide board deck. Three bullets and a visual per slide.
Rewrite these slides for a non-technical exec. Under 25 words per slide.
Generate two minute speaker notes per slide, UK English.
Docs
Tighten this proposal by 30 percent. Do not drop any commercial commitment.
Restructure. Give me a better TOC and a one paragraph rationale per section.
Fix to UK English. Single quotes, dd/mm/yyyy dates.
Draft a polite chase for this thread. One short paragraph, clear ask, UK English.
Summarise this 40 message thread into decisions, open questions, owners.
Triage unread, group by project, flag what needs a reply today.
Pattern. Role. Task. Constraints. Context. Format. Five beats. Every time.
What Anthropic does with our data.
Plain reading of the Privacy Centre and the Trust portal, as of this session. Worth knowing because it changes what is safe to paste.
No training on our content
Prompts and uploads are not used to train Anthropic models. That is the default, not a setting you have to toggle. Anthropic acts as our processor, IDWare is the controller.
You delete, it goes
Chats persist until a user or admin deletes them. Deleted content purges from backups within 30 days. Trust and Safety flagged content can sit up to two years.
US hosted by default
Claude.ai runs primarily in the US. No native UK or EU residency on Claude.ai Enterprise yet. If you need EU residency for a workload, route via Claude on AWS Bedrock in Frankfurt or Ireland.
Thumbs send the chat in
Clicking thumbs up or down on an answer opts that conversation in for review. Owners can disable feedback workspace-wide under Data and Privacy. Worth doing for sensitive Projects.
Treat it like a system, not a chatbot.
The moment Claude reads an email, a SharePoint doc, a webpage or a PDF, anything written inside that content can try to instruct it. That is prompt injection. With Connectors in play it is the real threat, not model hallucination.
Even on Enterprise
Customer PII or biometrics. Network diagrams with real IPs. Pen test outputs against named customers. Contract pricing. Partner IP under NDA. OFFICIAL SENSITIVE material. Sanitise to placeholders first.
Untrusted content drives the model
Booby trapped email. Poisoned web page. PDF with hidden white-on-white text. If Claude has tools, the attacker now has them too. OWASP LLM01, NCSC Secure AI guidance both name it.
Five habits that hold the line
Read only Connectors by default. Confirm every outbound action. No auto-run over shared inboxes. Log Connector activity weekly. Sandbox third party Skills with dummy data first.
Skills and MCP servers are extensions
Treat a community Skill or MCP server like a browser extension on a finance laptop. Identify publisher, read manifest, pin version, admin gate, allow-list reviewed quarterly.
Pick one thing.
The only way this turns into a habit is if you ship one thing out of this room. Three options. Each a 20 minute task.
Spin up a Project
Pick something you write more than once a month. RFP section, customer update, board report. Drop three good examples in, write the standing brief, run it once.
Write a Skill
Bottle a runbook you keep re-typing into Claude. Save it. Share it on the team SharePoint. Refine on the next use.
Enable a Connector
SharePoint via Graph MCP, or Google Drive if you live there. Read only first. One real task. See what changes.
QUESTIONS ?