| name | skill-evolution |
| description | Use before creating, editing, or deciding whether to update any AI coding agent skill in this repository, including corrections to existing skill behavior, references, or routing. |
| author | NVIDIA Corporation and Affiliates |
| license | Apache-2.0 |
Skill Evolution
Use this skill before creating, editing, or deciding whether to update any AI coding agent skill in this repository.
When to Update Skills
Update a skill when any of the following happens:
- A user corrects agent behavior in a way that should generalize.
- A command fails and the recovery path should be remembered.
- A reference link, command, component
_type, or installation instruction is stale.
- Task routing is too broad, too narrow, or points to the wrong skill.
- A repeated instruction appears in multiple skills and should move to a shared entry point.
- A skill contains too much detail and should move detail into focused references.
Do not update skills for one-off user preferences, temporary local environment quirks, or speculative guidance that was not validated.
Update Workflow
- Finish the user's requested task first unless the skill update is the task.
- Identify the smallest skill that should change.
- Keep each
SKILL.md concise and task-oriented.
- Move detailed examples or long reference material into that skill's
references/ directory.
- Update
AGENTS.md when a new skill is added or task routing changes.
- Run Markdown link checks through pre-commit on the changed skill files.
Naming and Layout
Use the flat skills layout:
skills/
nat-installation/
SKILL.md
references/
nat-workflow-creation/
SKILL.md
references/
Each skill folder must contain one SKILL.md with frontmatter:
---
name: skill-name
description: Use when ...
author: NVIDIA Corporation and Affiliates
license: Apache-2.0
---
Use author and license frontmatter instead of long license headers in SKILL.md files.
Use specific names that describe the task surface. Avoid catch-all folders that hide routing information.
Quality Bar
- Skills should tell agents what to do, what to read next, and what to validate.
- Prefer canonical repository docs over copied long-form explanations.
- Keep cross-skill links relative and valid.
- Use "NVIDIA NeMo Agent Toolkit" on first prose use, then "NeMo Agent Toolkit" or "the toolkit".
- Use
nat only for technical identifiers such as the CLI, package name, Python namespace, paths, and environment variables.
