Add Skills during agent creation
The current builder has five steps:
- Basics — name and describe the agent.
- Model — choose the provider and model policy.
- Tools & Skills — assign capabilities and procedures.
- Prompt — define the agent’s enduring role and limits.
- Review — inspect the complete configuration before saving.
In Tools & Skills, select one or more Skills. VDF compares each Skill’s declared tool requirements with the agent’s assigned tools and reports missing capabilities.
Use Add missing tools only after reviewing the gap. A declared capability might be unnecessary for your narrower use case, or it might require a connector the user should not receive.
Invoke a Skill explicitly
In the Agent Playground, type / to open Skill completion and choose an available Skill:
/option-comparison Compare these three deployment approaches against our criteria.Explicit invocation is the clearest test because it removes activation ambiguity while you evaluate the procedure itself.
Allow model-initiated activation
Compatible tool-calling providers can inspect Skill listings and load instructions when the current task matches a description. Provider behaviour differs, so test the exact model and configuration you intend to deploy.
Use explicit invocation when:
- The selected procedure is part of a controlled runbook.
- Two Skills have overlapping descriptions.
- A regulated workflow requires deterministic selection.
- The provider does not support the model-initiated loading loop.
Test four dimensions
| Dimension | Question |
|---|---|
| Trigger precision | Does the Skill activate only for relevant requests? |
| Trigger recall | Does it activate across the realistic phrasings users employ? |
| Output quality | Does the result follow the procedure and expected shape? |
| Safety | Does the agent stop, disclose gaps, and respect tool boundaries? |
Test positive examples, near misses, ambiguous requests, missing tools, unavailable sources, and a deliberately unsafe request.
Inspect events
A run can report:
- Skills available to the agent.
- Skills activated during the request.
- Bundled files read after activation.
- Listing and instruction context loaded.
- Approximate characters added to context.
Use these events to distinguish a triggering problem from a tool failure or a weak procedure.
Common failures
The Skill never activates
Make the description more specific, verify that the Skill is assigned, and test explicit invocation before changing the workflow.
The Skill loads but cannot act
Check the missing-tools warning, connector authorisation, and user permissions. Loading instructions cannot create access.
Too many Skills compete
Reduce overlapping assignments or sharpen descriptions. More available procedures can make selection less reliable and consume additional listing context.
Next step
Use versions, imports, and exports to preserve a known-good procedure.