run-<unit-name>

Create a SKILL.md file with frontmatter containing a name and description, then fill in sections for prerequisites, setup, build, run, and test commands. Include the exact commands that worked, not generic instructions, so the agent can execute them directly.

Write a driver script placed next to the SKILL.md file and launch it under xvfb for headless display. For REPL-style drivers, wrap the session in tmux and poll for a ready marker before sending commands and capturing the pane.

The description should contain action verbs like start, run, build, test, and screenshot, since those are the words an asking agent will type. The agent scans this description to decide whether to load the skill automatically.

Move the driver to the project's scripts/ or e2e/ directory when it grows into shared launch helpers or a real end-to-end test harness that the project's test suite needs. Update the paths in the SKILL.md afterward while the skill file itself stays in place.

Polling for a ready marker between tmux send-keys and capture-pane is faster than a fixed sleep and fails loudly instead of capturing a half-rendered screen. This makes agent-driven UI interaction more deterministic.