Prompts, Skills, and Playground

Fabric Experiments Studio separates three concerns that are often collapsed in an AI console:

This mirrors the document-oriented authoring experience in Fabric GTM Brain while preserving the Experiments boundary: Experiments stores and evaluates the instruction evidence; Fabric Harness executes agents; Fabric Platform governs business mutations.

Prompts and Skills use the same controlled master/detail shell, tree navigation, system-guide pattern, responsive behavior, and URL-addressable record selection from @fabricorg/ui. Product-specific forms, version evidence, evaluation behavior, and governed API actions remain owned by Experiments.

Markdown authoring

Prompt messages, Skill instructions, and model responses support GitHub-flavored Markdown. The shared editor provides Write and Preview modes plus controls for headings, emphasis, links, lists, quotes, inline code, and fenced code blocks. Rendered documents support:

• headings and nested lists,
• links and block quotes,
• tables and task lists,
• inline and fenced code,
• emphasis and horizontal rules.

Raw HTML is not executed. Prompt variables remain explicit {{variable}} placeholders and are rendered only when the Playground or evaluation executor runs the template.

Build a Skill

Open Build → Skills. The left navigator groups tenant-owned Skills by workspace or personal scope and folder. Select a Skill to review its rendered instructions, version evidence, test cases, permissions, tags, and approval policy.

Build a prompt

Open Build → Prompts. Prompt records and the versioned authoring guide share the same tree interaction as Skills. Creating a prompt happens in the detail pane rather than a disconnected full-page form. Selecting a record shows its immutable versions, hashes, notes, rendered Markdown messages, model parameters, and side-by-side diffs before authoring the next version.

Creating or revising a Skill captures:

1. stable Skill ID, name, description, scope, and folder;
2. Markdown instructions;
3. declared capability names;
4. whether human approval is required;
5. isolated test prompts and expected behavior;
6. tags and a version note.

Publishing unchanged executable content is idempotent. Changed instructions, permissions, approval policy, tests, or tags mint a new immutable version with a SHA-256 content hash. Lifecycle metadata can mark a Skill draft, active, or retired; Playground pins an explicit version even when viewing the latest head.

Skill permissions are declarations, not credentials or authorization grants. An agent tool that changes business state must still use the governed Fabric Platform action and its policy/approval boundary.

Compose in Playground

The Playground composer supports either a registered prompt version or an inline Markdown draft. Optionally select a Skill and version. The server composes the governed Skill instructions ahead of the prompt, renders variables, invokes the configured model endpoint, and returns a fully rendered Markdown response.

The result includes input/output tokens, latency, and endpoint identity. Every successful invocation writes a tenant-scoped audit event and, when the trace hot store is present, an LLM span containing prompt and Skill provenance.

Replay and evaluate

Replay against dataset pins:

• prompt ID and version,
• optional Skill ID and version,
• frozen dataset ID and version,
• evaluator versions and example limit.

Replay creates a durable evaluation run. Prompt and Skill composition happens in the run executor, model generation consumes the shared token budget, and evaluation findings flow into Evals and Quality Center. The request lifecycle does not perform the dataset-wide model loop.

API surface

Tenant-scoped REST resources are available under /v1/orgs/{orgId}:

GET, POST     /skills
GET, POST,
DELETE        /skills/{skillId}
GET, POST     /skills/{skillId}/versions
POST          /playground/invoke
POST          /playground/replay

Viewers may read Skills. Experimenters may create Skills, update metadata, publish versions, invoke Playground runs, and start replays. Deletion requires an admin or owner. API-key scope and organization membership checks still apply.

To build an executable custom agent that consumes these instructions, start with Fabric Harness's Your First Agent and Agent anatomy guides.