Draft architecture from evidence
Use an enabled AI client to turn source material into a small set of reviewable proposals.
On this page
Define the task before supplying material
Choose a result you can verify. For a fictional distributor, that could be the applications and data exchanges involved in order fulfilment. Confirm the workspace and the actions available in your MCP connection. Use source material your organisation has approved for the client.
Ask for a traceable draft
Review the supplied order fulfilment documents against the existing Atlas model. Draft the applications, capabilities, data objects and integrations the material supports. Give the source, any matching Atlas record and the reason for each proposed edit. List assumptions and open questions separately. Show me the draft before submitting anything.
Check that applications match the intended records, that each integration has the correct producer and consumer, and that source material supports any owner or lifecycle value. If the specification names an ERP but gives no owner, leave that responsibility open until someone can confirm it.
Validate and review the proposal
When available, use validate_changes on the draft. Validate changes and review model quality explains how to handle errors and warnings. Resolve material assumptions, then use the workspace’s proposal process. After implementation, read the affected records again and compare them with the reviewed proposal.
| Problem | Next step |
|---|---|
| The draft adds details missing from the source | Ask the client to remove those assumptions and list unanswered questions |
| Existing applications are proposed again | Supply the matching records and ask for a revised draft |
| A source describes a future design | Label the intended state and timing before review |
| The draft is too large to review | Split it by business area or decision |
Use Onboard an existing landscape to organise the wider effort.