At the start of every conversation, read the following in full to load all project rules and skills:
- All
SKILL.mdfiles anywhere underskills/. Use a traversal that includes every nested directory even when search tools would normally skip ignored paths, for example:find skills -type f -name SKILL.md -print
skills/ is the canonical, agent-agnostic location for SKILL.md-format workflow guides.
Cursor and Claude Code load the same skills via their respective pointer mechanisms (.cursor/rules/agent_skills.mdc, CLAUDE.md). For Codex (this file), the skills are background context — refer to them when the user's task overlaps a skill's subject.
| Skill | What it does |
|---|---|
isaac-sim-remote |
Drive a running Isaac Sim via the isaacsim.code_editor.python_server TCP socket (port 8226). Headless OK. |
profile-isaac-sim |
Profile and optimize with the in-repo benchmark scripts + Tracy; compare runs, diff frame times. |
validation-diff-gifs |
Pixel-diff GIFs of validation captures vs golden data; fastest path to triage benchmark image failures. |
Additional skills covering orchestrator + meta, URDF -> USD, physics, navigation/fleets, manipulation, sensors, rendering, USD pipeline, ROS 2 bridge, and operational meta-skills. They assume you are driving a built Isaac Sim from a script (SimulationApp, isaacsim.core.experimental.*, USD authoring, RL with Isaac Lab).
Do not load all at once. Use the categorized index in skills/SKILLS.md for pipeline diagrams, priority read order, the $ISAAC_SIM_DIR / $ISAAC_LAB_DIR / $WORKSPACE_DIR contract, and "when stuck" pointers.
.cursor/rules/ holds the *.mdc style and policy rules (C++ codestyle, Python codestyle, docs style, extension structure, build instructions, pip packaging, ...). Cursor loads them automatically; other platforms read them at session start.
Skills are authored once under skills/. .claude/skills/ contains symlinks into skills/ so Claude Code's native skill discovery works without duplication. .cursor/skills/ is retired.