Semogram Docs
Skills

Create with an agent

Review an AI-authored package before creating the workspace skill

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.

The creation agent turns a task description into a stored, reviewable skill artifact. Creating or revising that artifact does not yet create the standalone workspace skill. Validation and the final creation action are separate steps.

Describe a specific job

Open Skills → Create with agent. Describe when the guidance should apply, the inputs it needs, the output you expect and actions outside its scope. For this self-contained example, use a review procedure for the following fictional case:

Creation prompt
Create a standalone skill named equipment-risk-review for reviewing equipment
failure-risk predictions. Require equipment and prediction identifiers,
evidence cutoff, evaluation window, probability, timestamped observations and
observed outcome if available. Distinguish predictions from outcomes and flag
observations newer than the cutoff. End with missing information and next checks.

The bounded example is pump-17: probability 0.70, cutoff 2026-10-01T12:00:00Z,
window ending 2026-10-08T12:00:00Z, vibration observed 2026-10-02T09:00:00Z,
and no observed failure. The expected review says that the later observation
was not an original input and that 0.70 does not prove failure.

Do not create a forecaster, record an outcome or execute a maintenance action.
Show the package and validation before requesting creation.

Inspect the artifact

Read the exact SKILL.md revision, its name and description, proposed commit message, citations and validation issues. Check that the scope matches the job and that the agent has not described permissions or tool access that the workspace does not have.

The current generator produces a SKILL.md head. Add supporting files through the package editor when the job needs them; do not assume generation automatically builds a runnable script bundle.

Revise and validate

Ask for a focused revision when a check is missing. Each revision is a separate artifact revision; a valid earlier revision does not establish that the latest revision passed. Validation checks the package and whether a case-insensitive skill name already exists in the workspace.

If equipment-risk-review already exists, inspect and edit that skill or choose a deliberate different name. Creating a duplicate is not a version update.

Create the workspace asset

Review and approve the final creation action for the validated revision. The creation tool rechecks the draft before saving and links the artifact to the created skill ID and version. If agent creation is unavailable because the deployment's approval configuration is disabled, use manual instructions or upload; a generated draft is still not a saved skill.

Reopen the saved skill, inspect its head and compare a review of the supplied fictional case against the expected result. The first skill guide contains the full three-file package and exact acceptance table.

Input methods

MethodSupported behavior
Platform agentDraft, revise, validate and request approval to create
Manual editorCreate or edit text files and save a package revision
External assistant/MCPDo not assume skill creation is exposed as a public tool
Public HTTP APINo skill creation route is currently available