Customize Cloud Tag
Cloud Tag is customized per group, not per person. Most settings exist at several levels, and a group gets them from the organization, then its agent, then any bundles that apply to it, then itself:
| Setting | Organization | Agent (a linked mailbox) | Bundle | Group |
|---|---|---|---|---|
| Instructions | Organization-wide instructions | Instructions for @name, in every group | The bundle's instructions | Custom instructions |
| Model and effort | Default model and effort | Model and effort | — | Model and effort |
| Connectors, repositories, skills, secrets | On in every group | — | On in the bundle's groups | This group only |
| Auto-approve rules | On in every group | — | — | This group's own |
| Spend limit | Monthly limit, and default per group | — | — | This group's own limit (group admins up to the default) |
- Instructions add up. The agent reads the organization's, then the agent's, then each bundle's, then the group's, in that order. The group's page shows what it inherits under Instructions from …, added before these.
- Connectors and repositories take the narrowest for the same server or repository: the group's, then a bundle's, then the organization's.
- Model and effort take the narrowest one set. A group's own choice wins, then the agent's, then the organization's.
Instructions
Write a group's instructions on its Overview page, under Custom instructions. Say what the agent should know about the group: its purpose, tone, scope, who is who, and what not to do.
This is the support escalations group. Answer in under 150 words.
Billing questions go to Sam; never promise refunds.
Link the runbook in the group files when one applies.
Organization-wide instructions are on the organization page under Defaults for every group; an agent's are under that agent in 2. Agents.
Instructions are configuration, and the agent treats them that way. What people write in emails, attachments and tool results is information, not instructions, so a message can't override them.
When it replies
On the group's Overview page: When mentioned (the default) or Respond automatically. See Control when Cloud Tag responds.
Model and effort
On the group's Overview page, admins pick a model and effort, or leave Default to inherit. The organization's Owners and Admins set the agent's and the organization's. See Choose the model.
Skills
Skills are folders of instructions and scripts the agent loads when a task calls for them. On the Skills page:
- Built-in document skills: Excel spreadsheets (
xlsx), Word documents (docx), PowerPoint decks (pptx) and PDFs (pdf). All four are on for a new group; turn off the ones it doesn't need. - Custom skills: upload a
SKILL.md, or a.zipof a skill folder. Each can be turned off or removed. - Skill uploads must be under 4 MB. A zip can contain at most 256 files and folders, with no file expanding above 8 MB and no more than 32 MB expanded in total, including plugin content that isn't used. Encrypted zips and unsafe or duplicate paths are rejected.
- Plugins: upload a plugin's
.zip(or.plugin) to add all of its skills at once. It needs.claude-plugin/plugin.jsonat the top of the zip, or inside its one top-level folder, and each skill inskills/<name>/SKILL.md, the layout Claude Tag uses. Each skill is added separately, so you can turn off or remove them one by one. The plugin's commands, agents, hooks and MCP servers aren't used; add its servers as connectors. Plugins nested inside it aren't read either, norskillspaths itsplugin.jsonadds beyondskills/. To add a plugin from GitHub, zip its own folder: the repository's Download ZIP wraps it a level too deep.
Skills on the organization's Skills page are on in every group.
An agent attaches up to 500 unique skills, counting built-ins and inherited skills. Built-ins take priority, then the group's custom skills, bundles by name, and organization skills. The group's Skills page names any left out; turn off unneeded skills at their owning scope to free a place. Duplicate attached skill IDs count once, with the narrowest scope's version winning. Skills discovered in mounted repositories also count toward the provider's session limit, beyond this attached-skill selection.
Files
The Files page lists every attachment posted to the group, files admins added, and files the agent produced. The agent can search and open all of them from any thread.
To give it reference material, such as handbooks, schedules or price lists, upload a file there with a short description that helps it find the file. Each upload can be up to 4 MB.
Past mail
The agent knows only mail sent after it joined. Import earlier mail on the History page as an .mbox export. See Import past mail.
Memory
The Memory page shows the agent's notes for the group; admins can edit or delete them. On the Overview page, Let Cloud save to the notes it shares across all its groups from here controls whether this group can write to the agent's shared notes. See What it remembers.
Who can edit
Group admins and the organization's Owners and Admins can edit instructions and reply mode. Posting to the group does not grant web access or configuration rights. The former member-edit setting is unavailable, including for groups that previously enabled it. See Control access.
When changes apply
Changing instructions, the model, connectors, skills, secrets or network access re-makes the group's agent. Each thread moves to a fresh session on its next reply and reads the thread again, so nothing is lost and the change applies everywhere.
What stays fixed
- The agent posts only to its group, threaded under the message it answers.
- It never reveals what it knows from one group in another (shared notes, which admins control, are the exception).
- It confirms before consequential or irreversible actions with a connected service.