The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill performs dynamic execution of shell commands extracted from local documentation assets.
Evidence: In SKILL.md (Phase 1.5.4), the skill iterates through commands extracted from .tape files and executes them via bash -c "$cmd" to verify they produce expected output.
Evidence: The manifest parsing logic in modules/manifest-parsing.md defines a requires field that executes arbitrary command strings as background processes (e.g., npm run serve).
[COMMAND_EXECUTION]: The skill automatically detects build systems and executes commands to build or install binaries.
Evidence: It uses find, grep, sed, and awk to discover assets and parse commands.
Evidence: In SKILL.md (Phase 1.6), it detects Cargo or Make projects and executes cargo install or make build to ensure binary freshness.
Evidence: It executes the detected CLI binary with --help and --version flags during validation steps.
[INDIRECT_PROMPT_INJECTION]: The skill provides a significant attack surface for indirect prompt injection through local project files.
Ingestion points: Workspace files discovered via find, specifically .tape and .manifest.yaml files.
Boundary markers: None. The skill trusts and executes instructions found within these files as part of its validation pipeline.
Capability inventory: Full shell execution via bash -c, build system invocation (cargo, make), and background process spawning.
Sanitization: Absent. The extraction logic uses basic text filtering that does not prevent command injection or execution of malicious payloads embedded in the tutorials.
[EXTERNAL_DOWNLOADS]: The skill references external tools and packages for user installation.
Evidence: Instructs users to install vhs via Go and playwright via NPM from official repositories and registries to resolve missing dependency errors.

tutorial-updates