Configure a custom Expert
Before creating a custom Expert, check Template Experts. Template Experts are the fastest path when your workflow matches a common pattern such as PR authoring, code review, verification, feedback triage, incident investigation, project building, or ticket dispatch. Create a custom Expert when the workflow is specific to your team, requires custom tools or instructions, or combines systems in a way no Template Expert covers.1
Define the workflow
Write down what the Expert owns, what input it expects, what output it should produce, and where a human must make a decision. Good Experts have a narrow job and a clear stopping rule.
2
Choose the runtime environment
Select an Environment with the repos, CLIs, language toolchains, and network access the Expert needs.
3
Add integrations and capabilities
Grant only the integrations the Expert needs, such as code hosting, ticketing, communication, web access, or monitoring tools.
4
Write the system prompt
Describe the Expert’s role, inputs, process, output format, allowed write operations, and handoff rules. Include idempotency checks for event-driven Experts so repeated triggers do not create duplicate comments, tickets, or sessions.
5
Configure launch guidance
Add user instructions and placeholder text that tell people what to paste when launching the Expert, such as a change request, incident channel, ticket, project brief, or PR/MR URL.
6
Add workers or triggers if needed
Attach worker Experts for delegated sub-tasks, or add triggers for code-hosting events, communication messages, schedules, or webhooks. Test the manual path first, then automate.
7
Validate with a real example
Run the Expert on a representative task, inspect the session, and tune the prompt or integrations before rolling it out to the team.
Prompt checklist
Strong Expert prompts usually define:- Role and scope: what the Expert owns and what it must not do
- Required inputs: what to ask for before taking action
- Process: the phases it should follow, including when to stop for human input
- Output format: the exact shape of comments, reports, tickets, or summaries
- Allowed writes: which external actions are permitted
- Idempotency: how to detect prior runs and avoid duplicates
- Attribution: how messages link back to the Cosmos session that produced them
- Memory: where reusable team learnings should be read or written, if the workflow improves over time
Experts are saved configurations, not one-off prompts. Keep them small enough to reason about, test them on real examples, and prefer composing a few specialized Experts over one very broad Expert.