Create your first skill
Save a complete equipment review package and verify its use
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.
This example creates reusable equipment-risk review guidance. You do not need a database or live forecaster to complete the package and review test: the fictional case below contains every input. It tests the review procedure, not a real prediction pipeline.
Create the instructions
Open Skills in a project. From the creation menu choose Write skill instructions. The editor opens an unsaved package whose head is SKILL.md. Replace its head content with:
---
name: equipment-risk-review
description: Review equipment failure-risk predictions against dated evidence and a defined evaluation window.
---
# Equipment risk review
Use this guidance when reviewing an equipment failure-risk prediction.
1. Identify the equipment ID, prediction ID and the question being answered.
2. Record the evidence cutoff and evaluation window before interpreting risk.
3. Compare each observation timestamp with the cutoff. Flag observations newer
than the cutoff; do not present them as inputs to the original prediction.
4. Separate the predicted probability from an observed equipment failure.
5. Report missing evidence and the effect on what can be concluded.
6. Follow references/review-checklist.md and use references/example-case.md
as the bounded comparison fixture.
7. End with: identifiers, evidence timing, prediction versus observation,
missing information, and a proposed next check.
Do not change the prediction, record an outcome, approve a write or claim that
maintenance was performed. Those actions require their own supported workflow.Use the file tree to add these two supporting files:
# Review checklist
- Equipment, prediction and saved forecaster revision are identified.
- Evidence cutoff and evaluation window are recorded.
- Observations after the cutoff are distinguished from original inputs.
- Predicted probability is not treated as a confirmed failure.
- Missing evidence and next checks are explicit.# Example case
This is a fictional review fixture, not a real prediction record.
Equipment: pump-17
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
Predicted probability of failure: 0.70
Observation: vibration measurement at 2026-10-02T09:00:00Z
Observed failure: not supplied
Expected review: the vibration observation is after the evidence cutoff;
it was not an input to the original prediction. A probability of 0.70 does
not establish that a failure occurred. Request evidence available at the cutoff
and an observed outcome for the defined evaluation window.The paths in the instructions refer to files inside this package. They are not a data endpoint, ontology term or prediction ID. The fixture identifiers are fictional labels and must not be passed to a live lookup as UUIDs.
Validate and save
Check the editor's validation summary, frontmatter, nonempty instructions and each referenced file. In Save, enter a commit message such as Create equipment review guidance, then choose Create skill. Reopen the installed skill and confirm its files, asset ID and saved version.
The workspace now contains a reusable skill. No prediction has run and no maintenance outcome has been recorded.
Run a bounded review
Open the supported authoring/review conversation for your task. Provide the package instructions and fixture content, or use a context that explicitly loads them. The list of available skills supplies names and descriptions; do not assume a name alone loads the full package.
Use the equipment-risk-review instructions and its two reference files supplied
in this conversation. Review the fictional pump-17 case. State the evidence
cutoff, explain whether the later vibration observation was an original input,
and separate failure probability from a confirmed outcome. Propose next checks
without changing a prediction or recording an outcome.Check the result
| Expected result | Why |
|---|---|
| Cutoff is 1 October 2026 at 12:00 UTC | The case states the evidence boundary |
| The 2 October observation is after the cutoff | It cannot be presented as an original input |
| 0.70 is a predicted probability | No observed failure is supplied |
| Outcome and cutoff-valid evidence are requested | Those are missing inputs needed for review |
| No change is claimed | The task asks for a review only |
Keep the reviewed result and skill version as the baseline. If it fails these checks, inspect what instructions were actually supplied before editing the skill. Saving a package alone is not evidence that a model consumed it.
Download instead of typing
Download the complete ZIP, SKILL.md, checklist, fixture or serialized package. The upload guide explains each input format.