The Agent Skills Directory

[COMMAND_EXECUTION]: The skill frequently executes local shell scripts located in ~/.claude/skills/nanostack/bin/ to manage project state, handle session transitions, and generate reports. These include init-project.sh, session.sh, resolve.sh, save-artifact.sh, and sprint-metrics.sh.
[DYNAMIC_EXECUTION]: The skill dynamically loads and executes shell scripts using the source command (.) from a computed path in the user's home directory (~/.claude/skills/nanostack/bin/lib/skill-preamble.sh and skill-finalize.sh).
[EXTERNAL_DOWNLOADS]: The search phase described in references/search-before-building.md allows the agent to fetch package information and README content from public registries including npm, PyPI, and GitHub. These are well-known services and the skill uses them to identify existing solutions.
[DATA_EXFILTRATION]: The skill contains an opt-in telemetry feature that transmits metadata such as skill duration, execution outcome, and system architecture to a remote server. It provides clear disclosure that no sensitive code, prompts, or personal data are included in these transmissions.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data from web searches and project artifacts, creating a potential surface for indirect injection attacks.
Ingestion points: External data from npm, PyPI, and GitHub search results; local sprint journals, session JSON files, and metrics reports.
Boundary markers: The skill contains high-quality instructions in references/search-before-building.md defining a "Prompt-injection boundary" to ensure fetched data is treated strictly as information rather than instructions.
Capability inventory: The skill possesses the capability to execute shell commands, perform network requests via curl and the gh CLI, and read/write to the local filesystem.
Sanitization: Instructions explicitly direct the agent to ignore directives, role-play attempts, or commands found within external content.

think