Manage AI skills in the Playground

Every AI capability in DeltaVault is a published skill you manage in the AI Playground, in the admin settings. This is the skill-centric editor: a rail of skills on the left, an editor in the middle, and a test runner that stays visible so you can edit and test without leaving the page. This guide covers the parts a platform administrator works with.

The rail, grouped by surface

The rail lists every skill grouped by the surface it appears on (Business Modeling, Catalog, Map and Match, Settings Profiles, and so on), so you can see at a glance what runs where. Hover any item to read its description, the same one-line summary users see under the skill.

The System AI group

Skills that power built-in behavior (the Ask AI chat, the workshop, discovery, consolidation, governance checks, document classification, and the Model Review passes) are collected under a System AI group, each marked with a System badge. These used to be prompts hardcoded in the product. They are now published skills like any other, so you can read and tune the exact instructions the product runs.

Editing a built-in skill is guarded

Because a built-in skill has to keep the shape the product expects, its structural fields (its inputs, which actions it may take, and how results come back) are read-only. You can still edit what matters for tuning: the system prompt, the Default effort, and the token budget. A note on the editor spells out which fields are locked and why, so nothing changes silently.

Default effort replaces the model dropdown

A skill no longer names a specific model. Instead, the Settings tab has a Default effort selector with four settings, Quick, Standard, Deep, and Expert, and shows the model each setting maps to as helper text. Effort is the default a run starts from when the person running it has not chosen their own; they can still override it per run from the surface where they launch the skill.

Context layers and the Conventions toggles

Every skill declares which context it reads. The Context tab lists the layers, including a Conventions group with three checkboxes: Categories, Naming convention, and Classifications. Each checkbox injects the matching section of Settings → Conventions into the skill’s prompt: the category catalog with each lane’s nature and guidance, the entity and attribute name casing, or the classification schemes and their allowed values.

Prompt authors can position the injected text precisely with the {{conventions}} placeholder. When the placeholder is absent, the rendered block is appended to the end of the system prompt automatically, so the toggle alone is enough for most skills.

The test runner shows the exact rendered text the model receives. Run a preview and expand the context blocks: each layer appears on its own, so you can read the categories, naming, and classification text word for word before publishing a change.

The test runner

The test panel runs the skill you are editing without leaving the Playground. For a built-in skill, it renders the prompt’s expected placeholders as plain text inputs so you can supply sample values and see the result. The edit-test-edit loop stays on one page.

Orchestrated skills point to their prompts

Some launcher tiles (for example the knowledge consolidation and governance-coverage tiles, and Review business model) orchestrate several system skills rather than carrying one prompt of their own. Their Prompt tab shows a short note explaining this and links to the system skills whose prompts actually run, so you always know where to edit the real instructions.

Workshop traces in Run History

A business modeling workshop is an orchestrated operation: its discovery phase runs several skill calls in sequence (finding events, gathering concepts, and mapping relationships) instead of one. Each of those discovery steps shows up in Run History like any other skill run, with an Operation badge that reads “Workshop” followed by the last six characters of the workshop’s own identifier (the same identifier appears on every step of one workshop), and the badge links back to that workshop. Use the Operation filter to isolate every discovery step that belongs to one workshop, instead of scrolling past unrelated runs.

You do not need to leave the workshop to see this detail. On the workshop run screen itself, a platform administrator can expand any discovery step in the live progress feed to see the model it ran on, the tokens it used, how long it took, and how many tool calls it made. From there you can jump straight to the full run record in Run History, or open the same step in the Playground to replay it against an edited draft skill.