This is a redacted copy of an internal document used at Coefficient Giving. The original document was authored by Chris Webster.
This document is shared in the hope it'll be a useful example for other groups in the community. Coefficient Giving takes no liability for any of the advice here. Note that many tradeoffs we've made (e.g. on security posture, approved apps) may pencil out different for your team.
This copy was made in early January 2026. It will not be updated, so if you're reading this more than a few weeks from now, parts of it may be out of date!
Please share your feedback! What do you do differently, and why? What new things should we try? Mistakes we're making? Please comment on this document or email Chris.
<aside> ✅
<aside> 🌐
<aside> 🚀
Claude Code is better for coding and is more feature rich. But if you prefer a visual interface over the Terminal, Cowork is a great option, and it's much easier to set up. Here's how they compare:
| **Claude.ai (Chat)** | Cowork | Claude Code | |
|---|---|---|---|
| Interface | Web browser or desktop app Chat tab | Visual desktop app (Cowork tab) | Terminal / command line |
| Best for | Quick Q&A, writing and editing, brainstorming, analyzing uploaded files, generating artifacts (code snippets, documents, diagrams) | File organization, creating Office docs (Excel, Word, PPT), browser automation via Chrome extension, research tasks, long sets of tasks | Building custom tools and automations, running shell commands, working with Git/GitHub, complex multi-file tasks |
| Integrations | Connectors (Notion, Slack, Google, Asana) — same as Cowork | Built-in connectors (Notion, Slack, Google, Asana) — toggle on in settings | MCP servers — more powerful but requires tricky setup |
| Multi-step tasks | No — single-turn responses to your prompts | Yes — shows a plan, you approve, it runs | Yes — plus Plan Mode, subagents, and parallel sub-tasks for more complex work |
| File handling | Upload files to chat; can generate files as artifacts to download | Reads and writes directly to a folder on your computer — real files you can open immediately | Full filesystem access in your project directory — reads, writes, creates, and deletes files |
| Memory | Memory across sessions and chat search to find past conversations | No memory between sessions | CLAUDE.md files persist context across sessions |
| Customization | Projects with custom instructions and knowledge files; personal styles | Plugins (Anthropic ships 11) | Subagents, skills, slash commands, and CLAUDE.md — highly configurable |
| Learning curve | Lowest — just type and go | Low — point-and-click, no terminal needed | Medium — requires comfort with the command line |
| Maturity | Most mature and stable | Research preview — can be buggy | More mature and stable |
<aside> ⚠️
<aside> ⚡
Some things you can ask:
Cowork is great at working with Office docs:
Example prompts that work well:
"Here's a CSV of survey responses. Analyze it and tell me the most interesting results."
"Here's my notes on using AI. Make me a 15-minute powerpoint for our staff."
You describe what you want, Claude breaks it into steps, you approve, and it runs. Sometimes for several minutes.
Examples:
<aside> 🔌
Connectors let Claude talk to other services. Cowork reads and writes local files, so you can pull data from Notion into a spreadsheet, or search Slack and save the summary to your folder.
Go to claude.ai/settings/connectors, turn on what you want, grant permissions. They'll show up in Cowork automatically.
</aside>
<aside> 🧩
Plugins bundle skills, connectors, and tools for specific jobs. Anthropic ships 11:
Browse them at github.com/anthropics/knowledge-work-plugins. You can customize them or build new ones.
Plugins save locally for now. Org-wide sharing is coming.
</aside>
</aside>
<aside> 🌐
</aside>
<aside> 📚
</aside>