Use and evaluate

You need a Semogram account with membership in the target workspace and access to a project in it. The Skills screen is opened from a project, while standalone skills belong to the workspace.

Use a skill for repeatable judgment or authoring procedure: how to check a draft, inspect evidence or organize a review. Use a transform or executable tool for deterministic data operations. A package's instructions can describe a tool but do not install it or grant permission to call it.

Current consumption boundary

Studio's authoring context lists workspace skill IDs, names and descriptions. That listing identifies available guidance; it is not proof that the full SKILL.md and every supporting file entered the model context.

The Studio skill configuration declares skillIds and the following assist modes:

These mode values are configuration, not a promise that a skill node independently runs. The skill node has no default execution tool in the graph executor. Before relying on a skill's effect, verify the consumer loads the package; otherwise provide its relevant text explicitly in the authoring/review conversation.

Example: review a prediction

Use this complete fictional input, plus the instructions below:

Review pump-17's failure-risk prediction using this procedure:
record identifiers, evidence cutoff and evaluation window; compare observation
times with the cutoff; separate probability from observed failure; report
missing information and a next check. Do not modify records.

Prediction: fixture-prediction-001
Forecaster revision: fixture-revision-001
Evidence cutoff: 2026-10-01T12:00:00Z
Evaluation window: 2026-10-01T12:00:00Z to 2026-10-08T12:00:00Z
Failure probability: 0.70
Vibration observation: 2026-10-02T09:00:00Z
Observed failure: not supplied

The expected review says the later observation was not an original prediction input, 0.70 is not proof of failure, and outcome/cutoff-valid evidence is missing. This task reviews supplied material and requires no live forecast or database endpoint.

Evaluate a revision

1. Retain a known input and the old review result.
2. Record the saved skill version and which files you actually supplied.
3. Change one instruction deliberately and save the reviewed revision.
4. Repeat the same task and compare the relevant conclusion, missing information and evidence treatment.
5. Check a second case so the instructions do not merely repeat the first answer.

For that second case, move the observation to 2026-10-01T09:00:00Z. The review should now recognize it as before the cutoff, while still distinguishing probability from an unreported outcome.

Keep lifecycle actions separate

Reviewing a prediction does not record an observed outcome. Writing skill instructions does not validate a live connector. Activating a revision does not repair an old run. Use the corresponding feature action when the review identifies a real change that needs to be made.

Package reference describes file contracts; troubleshooting helps separate missing context from invalid content or permissions.