Agent Skills
NeMo Gym ships agent skills — curated instruction files that teach AI coding assistants how to work effectively in this repository. They encode endorsed workflows for common tasks like adding benchmarks, debugging rollouts, and maintaining the docs site, so agents can follow project conventions without guesswork.
Skills are versioned with the repo and kept up to date as Gym evolves. Gym will continue to add skills as new workflows stabilize.
Where Skills Live
The .claude/skills/ tree is the superset — it includes Gym-specific skills plus vendored skills such as gh-stack for stacked pull requests.
Assistants discover skills automatically when they are pointed at a local Gym checkout. No extra install step is required beyond cloning the repository.
Available Skills
The Codex tree currently includes the five nemo-gym-* operational skills (excluding add-benchmark, nemo-gym-docs, and gh-stack).
Using Skills
When you work with an AI coding assistant in a Gym checkout, mention the task in natural language — for example, “add a new benchmark” or “debug this rollout failure.” Assistants that support skills will match your request against the skill descriptions and follow the bundled workflow.
You can also point an assistant at a specific skill file if you want a particular playbook:
The nemo-gym-docs skill mirrors the procedures in the Fern docs README — use it when contributing documentation changes.
Related Topics
- Development Setup — environment setup, CI, and commit requirements
- Add a Benchmark — human-readable tutorial that complements the
add-benchmarkskill - Contribute Environments — guidance for new training environments