Write reusable Agent instructions
An Agent page stores reusable instructions, skills, personas, or memory. It is content that an AI client can read; creating one does not start an autonomous worker or make every external client obey it.
For the Product Launch exercise, an instruction could say:
When drafting launch material:
- Use plain language for a new team member.
- Cite the workspace sources for factual statements.
- List unresolved questions rather than inventing answers.
- Submit proposed Context edits for human review.
Choose a category
| Category | Intended content |
|---|---|
| Agent guide | Shared working rules and guardrails. |
| Persona | The role and communication style to use. |
| Skill | Reusable task instructions, such as checking a release brief. |
| Memory | Background knowledge to carry into related tasks. |
Sensitive categories and sensitivity settings restrict visibility and activation. Avoid putting passwords or service keys into an instruction page.
Create, review, and activate
- An Editor or Owner opens Agent and chooses the relevant Create … Agent page or new-page action.
- Enter a title, category, body, and reason. Use a template only after reading and adapting it to your workspace.
- Choose Create draft. A saved draft is not active.
- An Owner checks the content and any required review, then uses Activate draft when permitted.
- Confirm the reader shows the intended active version.
Activation in the ordinary workspace UI requires the workspace Owner and cannot be performed by MCP/agent actors. Sensitive pages also require the workspace Owner. An Editor can author a proposal without being allowed to activate it. Being a server administrator does not by itself grant access to workspace content.
Editing an active page creates a pending draft. The existing active version remains in use until the replacement is activated. Restoring a historical Agent version creates a new draft; it does not immediately change the active version.
Make the client use the page
The client must explicitly read the Agent page or request an Agent input bundle through the integration. Merely storing or activating a page does not inject it into every prompt. The current Ask answer composer should not be assumed to apply all Agent pages automatically.
A bundle is a selected, version-pinned package of readable instructions and requested evidence. It has a token budget and an expiry. The Bundle tab can show records that included or omitted the page.
Selection is rule-based in this version:
- Always and Default on policies can include readable active pages.
- A requested category can include matching pages.
- Task matched compares terms from category, title, and slug against task intent; it is not a language model deciding the best skill.
- Permissions, disabled policy, draft status, and token budget can omit pages.
The word “Always” describes bundle selection after permission checks, not an ability to override permissions or an assurance that an external client consumes the bundle.
If an instruction seems ignored
Check the active version, whether the client read it or built a bundle, and whether the page was omitted. Use Chats to inspect recorded tool calls. If the bundle expired, build a new one. The external client remains responsible for using the returned content in its workflow.
Next: MCP connections and Review.
Implementation reference: Agent UI, activation policy, bundle selection.