Understanding Skills from a developer's perspective
Who this is for
Developers writing Skills.
A Skill written by a developer is an SOP the Agent can invoke — not a wiki written for humans.
Step 1. Define the runtime boundary
State explicitly in the Skill body:
- Which CLI commands / dataset codes / SQL codes / BFF names may be invoked
- Which steps must run with
dry-runor require human confirmation (see rabetbaseriskLevel)
Step 2. Map steps to platform assets
| Skill step | Platform asset |
|---|---|
| Inspect structure | rabetbase dataset detail |
| Query data | lovrabet data filter |
| Complex statistics | custom SQL code |
| Write-back | Backend Function |
Step 3. Versioning and naming
- skillCode is globally unique; a convention such as
team_scenario_v1is recommended. - Bump the version on breaking changes and keep the old Skill read-only.
Step 4. Integration testing
- Save the Markdown under Skills.
- Do a test run in a conversation using the skillCode.
- If it fails, revise the Skill or add to the Knowledge.
References
- Skill development guide
- The docs library entry on Rabetbase CLI risk levels