Guide1 min readUpdated October 6, 2026
How to document AI systems in a YAML or JSON manifest
Short answer
Give each AI system a distinct name, record its purpose/provider and declared controls, and keep unknown values explicit. Export YAML or JSON for review and version control. This manifest describes settings; it does not execute a policy or verify production behavior.
- Describe each system separatelyAdd a system for each distinct feature you want to review. For example, a support chat and an image generator can have different providers, purposes and human-review processes. Stable local IDs remain attached as you edit.
- Record what you actually knowUse Yes, No or Unknown for interaction, disclosure, generated content, human review and sensitive decisions. Unknown exports as null instead of false. Optional provider and purpose fields also remain explicit when unspecified.
- Record retention without guessingEnter the real declared application retention in whole days or leave it blank. Zero means zero days; blank means unknown. The generator does not recommend a legal retention period or inspect provider-side data settings.
- Review and version the exportUse the readable summary, choose YAML or JSON, then copy or download. Review names, versions and potentially confidential details before committing to a repository. Revisit the document when runtime settings or providers change.
Try it: AI System Manifest GeneratorRecord an application’s AI systems, declared controls and retention settings in a versioned local working manifest.Open tool
A working format needs a documented consumer
The schema identifier is utilumo-ai-system-manifest and version is 1. Systems are an array with distinct names. A downstream tool would need to explicitly support this schema to consume it. The file is not an official AI Act standard, executable compliance policy, audit certificate or a substitute for technical evidence.
Local inputs, deliberate exportsInputs and results stay in this browser tab. Download the record you need; review confidential details before sharing an export.