skills

A monorepo of agent skills for Unreal Engine + AngelScript workflows, following the skills.sh convention.

A skill is any directory under skills/ containing a SKILL.md file. The directory name is the skill identifier (kebab-case). Skills are agent-neutral — anything that speaks the skills.sh layout (Claude Code, Codex, Cursor, etc.) can install and invoke them.

Per-project pick: on any UE 5.x project, install one of unreal-engine or unreal-engine-angelscript (depending on whether the project uses the Hazelight fork). Pair with read-ue-logs always, and with ue-angelscript-tests if you're in an AngelScript project.

The two testing skills above are designed to pair with an MCP server that runs the headless test loop for the agent, so it never hand-types UnrealEditor-Cmd … Automation RunTests. Two options:

Our own focused Go server, scoped to the headless test loop. It exposes exactly five tools — status, list_tests, run_tests (headless -nullrhi), run_visual_tests (GPU), get_test_log — and nothing else. Install with go install github.com/osseous/ue-headless-mcp/cmd/ue-headless-mcp@latest and register it in .mcp.json (env: UE_EDITOR_PATH, MCP_UNREAL_PROJECT, optional UEHM_TIMEOUT_MS).

It exists because the editor process never reliably exits on Windows: it detects completion from the run log (the **** TEST COMPLETE. EXIT CODE: N **** marker, the AngelScript Hot reload failed due to script compile errors marker, and the optional -ReportExportPath index.json), sends child output to the null device to avoid inherited-pipe deadlock, and force-kills the whole process tree via a Windows Job Object the instant results exist. A clean run returns in ~20–45s; a compile error returns in ~20s with the extracted AngelScript errors instead of hanging. Each run writes a dedicated Saved/Logs/McpTest_*.log that read-ue-logs / get_test_log can read.

It does not build, cook, or drive the live editor (call_function, spawn_actor, pie_control, capture_viewport, console commands). Build C++ via UBT (Build.bat) directly; add the server below side-by-side if you need live editor control.

A broader Go server that also exposes build_project / cook_project and live editor control via the Remote Control API (:30010) and the MCPUnreal editor plugin (:8090). It tied "done" to the editor process exiting and never force-killed the Windows process tree, so a run whose tests finished in seconds blocked until the client timeout — which is why ue-headless-mcp replaced it for the test loop. Re-add it side-by-side only when you genuinely need its live-editor / build / cook tools.

The skills still work without any MCP server — they fall back to the Session Frontend / Automation RunTests CLI — but a project that mandates the MCP path (in its CLAUDE.md) should keep all test execution on the server.

npx skills add osseous/skills/<skill-name>

The installer drops the skill at .claude/skills/<skill-name>/ (for Claude Code) or .agents/skills/<skill-name>/ (agent-neutral).

Clone the repo and symlink everything into ~/.claude/skills/:

git clone https://github.com/osseous/skills.git
cd skills
bash scripts/link-skills.sh

Override the install destination with CLAUDE_SKILLS_DIR:

CLAUDE_SKILLS_DIR=/path/to/agent/skills bash scripts/link-skills.sh

On Windows, symlink creation requires Administrator rights or Developer Mode (Settings > Privacy & security > For developers). Without it the script falls back to a directory copy.

skills/                          # this repo
├── LICENSE                      # MIT, applies to every skill
├── README.md                    # this file
├── scripts/
│   ├── list-skills.sh           # prints every skills/**/SKILL.md path
│   └── link-skills.sh           # symlinks each skill into ~/.claude/skills/
└── skills/                      # all skills live here
    ├── unreal-engine/
    │   ├── SKILL.md             # frontmatter + agent entry point
    │   ├── README.md            # human-facing docs
    │   ├── references/          # spillover: cpp-style, replication, gas, lyra, ...
    │   └── scripts/             # detect-engine.ps1, find-uclass.ps1, open-epic-docs.ps1
    ├── unreal-engine-angelscript/
    │   ├── SKILL.md
    │   ├── README.md
    │   ├── references/          # cpp-differences, replication, mixins, footguns, ...
    │   └── scripts/             # detect-angelscript.ps1, grep-binding.ps1, open-as-docs.ps1
    ├── read-ue-logs/
    │   ├── SKILL.md
    │   ├── README.md
    │   └── scripts/             # read-logs.ps1
    └── ue-angelscript-tests/
        ├── SKILL.md
        ├── REFERENCE.md         # full API surface
        └── EXAMPLES.md          # copy-pasteable test scaffolds

1. Create skills/<your-skill-name>/SKILL.md with YAML frontmatter:

   ---
   name: your-skill-name
   description: <capability>. Use when <specific trigger contexts>.
   ---

2. Keep SKILL.md short (target < 100 lines). Spill detail into sibling REFERENCE.md / EXAMPLES.md.
3. Put deterministic helpers (scripts, templates) in <your-skill-name>/scripts/.
4. Run bash scripts/list-skills.sh to confirm the new skill is discovered.
5. Run bash scripts/link-skills.sh to install it locally for testing.

The description field is the agent's only routing signal — lead with the capability, then include explicit "Use when…" triggers so agents pick the right skill for the right task.

MIT © 2026 Maxim Kostin