Skills
Using and managing AI skills, including grounded-reply skills.
Skills are purpose-specific prompt packages the AI can load mid-conversation to produce higher-quality, opinionated output. When a skill is loaded, its guidance is layered into the system prompt for the rest of the turn and takes precedence over the generic response rules. Skills stack — the AI can load more than one and their guidance composes.
The AI loads a skill automatically when your request matches one; you'll see a chip such as Skill: Client Proposal in the conversation. You don't need to invoke skills by name.
Document skills
These produce a single, polished document rendered in the Hub as an artifact:
| Skill | Produces |
|---|---|
| Executive Report | A structured executive report with typography and KPI framing |
| Project Status | A status document with KPI cards and a work grid |
| Client Proposal | A client-ready proposal: cover, approach, deliverables, timeline, pricing, acceptance |
| One-Pager | A single-page briefing, memo, or decision document |
| Report | A generic, cleanly-formatted document |
Grounded-reply skills
These skills answer a request by retrieving context first, then responding — they do not answer from the model's general knowledge alone. When one is active, the AI reasons about what information the answer depends on, searches the relevant connected tools and internal artifacts, and grounds its reply in what comes back.
| Skill | What it does | Requires |
|---|---|---|
| Email Reply | Reads an incoming email, retrieves the context it needs (prior conversations, product docs, project status), and composes a draft reply in your Drafts folder | Microsoft 365 connected |
| Sales Reply | Answers a prospect/customer question (advantages, comparisons, objections) grounded in product, marketing, and proposal material | — |
| Knowledge-Grounded Q&A | Answers a factual question about your organization strictly from retrieved sources, and names those sources | — |
The Email Reply skill creates a draft only — it never sends mail on your behalf. Review and send the draft yourself from Outlook.
Grounded-reply skills are only as good as what's connected. If the relevant integration isn't connected — or a connector's search returns little — the AI will tell you what it couldn't verify rather than fabricate specifics. Cosmo does not yet index internal artifacts or memory for semantic search, so internal lookups rely on each connector's own search and exact/keyword matches.
Retrieval-first behavior
Whenever you have connectors or internal artifacts available, the AI follows a retrieval-first contract for any request that depends on facts about your tools, people, or documents:
- Name what it needs — the facts the answer depends on.
- Decide the sources — which connected tools would hold them.
- Fan out and retrieve — search and read across those tools.
- Ground the answer — synthesize only from what was retrieved.
- Flag thin grounding — say what's missing instead of inventing detail.
This applies to all chat, not just the grounded-reply skills; the skills add task-specific structure on top of it.
Availability
Skills that depend on a tool are hidden until that tool is connected — for example, Email Reply only appears once Microsoft 365 is connected. See Integrations to connect your tools.
Team skills
Besides the built-in skills above, a team can author its own. A team skill is shared across every workspace on that team — there's nothing to duplicate or move when you create it once.
Open Skills in a team's section of the Hub sidebar to see that team's list. Any member of the team can see and use its skills; creating, uploading, editing, activating/deactivating, and deleting one is limited to the team's owner.
For each skill you set:
- Name and a one-line description.
- Triggers — comma-separated keywords the AI checks your prompt against to decide when the skill applies, the same way it matches built-in skills.
- Instructions — what the AI should do when the skill is loaded.
Uploading a skill
If you already have a skill written — for example a Claude-format SKILL.md — the team owner can bring it in instead of retyping it: Add Skill → Upload a skill, then drag files onto the drop zone or browse for them. You can add several at once.
| File | What it must contain |
|---|---|
.md | YAML frontmatter at the top with name and description, followed by the instructions |
.zip or .skill | A SKILL.md (at the root, or inside a single top-level folder). Any other files — scripts, references, assets — are stored with the skill |
---
name: proposal-writer
description: Drafts tender proposals using the approved company template.
---
# Proposal Writer
Instructions for the skill go here...nameuses lowercase letters, numbers and hyphens (up to 64 characters);descriptionis up to 1024 characters; the instructions can't be empty.- Optionally add
triggers— the same keywords as the Triggers field, as a list (triggers: [review this email, check email tone]) or one comma-separated line. Claude's own format has no triggers, so without them the skill is matched on its description alone; adding a few makes it much more likely to be picked up in chat. You can also add them afterwards by editing the skill. - Each file can be up to 10 MB.
- Each file is checked on its own: a file with a problem is listed with the reason, and the rest still upload.
- If the team already has a skill with that name, you're asked whether to replace it or keep the existing one — nothing is overwritten silently. The same name on another team is fine.
- Files in an archive are stored as data. Nothing in them is ever run.
An uploaded skill is an ordinary team skill afterwards: you can edit, deactivate or delete it like any other.
Downloading a skill
Any member of the team can download a skill with the download button on its row — to back it up, share it with another team, or use it in Claude. A skill downloads in the same format it uploads in: <name>.md (with its name, description, triggers and instructions), or <name>.zip when it has supporting files. The file can be uploaded again as-is. A skill created by hand with a free-text name (e.g. "Executive Report") downloads with a lowercase-and-hyphens name (executive-report), since that is the form uploads and Claude accept.
A team skill is available in every new chat, not just while you're looking at that team — there's no need to open the team's Hub, Skills page, or a workspace on it first. This is checked once, when the chat starts: being added to a team, or a skill changing, won't update a conversation you already have open — start a new chat to pick up the change.
You can also load a specific skill yourself by typing @ in the composer — every skill you have access to shows up there, labeled with its team, and picking one tells the AI to use it for that message.
Toggling a skill to Inactive removes it from availability immediately for any new chat; it does not retroactively change a conversation already in progress.