Run the real application

The goal is not merely that a command exits zero. The goal is a live handle on the surface a user or downstream program actually touches.

Find the project recipe first

Before inventing commands, search the relevant repo/package scope for project skills. Prefer the most specific matching run-* recipe for the files or package being worked on.

If no usable project recipe exists, inspect the smallest set of authoritative launch hints: package/build manifests, README/developer docs, container config, executable entrypoints, and server config. If the launch path is unusual or took real discovery, use run-skill-generator after you get it working.

Identify the runtime surface

Product shape	Runtime surface
Web/desktop GUI	pixels plus browser/window interaction
CLI	command line, stdout/stderr, exit status
TUI/REPL	interactive terminal session
HTTP/API/server	listening socket plus real request/response
Library/SDK	public package boundary from a consumer process
Agent/prompt	real agent invocation and resulting behavior
Workflow/CI	actual workflow run

Do not substitute an internal function call for the public surface when a public surface exists.

Launch safely

Use the project's documented install/build/launch steps. Reuse existing dependencies and credentials; do not invent secrets. Isolate shared state with ephemeral ports, temporary directories, named tmux sessions, and test/sandbox accounts where practical.

Never perform destructive external writes just to prove the app starts. Use a dry run, sandbox, disposable target, or stop at the risky boundary and report it.

Drive the smallest meaningful flow

Exercise the smallest path that reaches the intended behavior:

• UI change -> navigate to the changed screen and interact with it.
• CLI flag -> invoke the shipped/public CLI with the flag.
• API handler -> send a real request through the listening server.
• error handling -> trigger the actual error from the public surface.
• persistence -> perform the action, reload/reopen, and observe persisted state.

Capture evidence as you go: screenshot, pane capture, stdout/stderr, response body, or another runtime artifact.

Report

Return the launch method, exact surface driven, observed result, evidence, any environment limitation, and cleanup status. Tests/typecheck are not proof that the application itself ran.

See references/surfaces.md when the project has multiple possible runtime surfaces.