The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes shell and Python scripts to automate project setup. It installs persistent lifecycle hooks into the ".claude/settings.json" file. These hooks execute local bash scripts from the skill's installation directory to perform tasks like updating the project wiki and tracking task focus.
[EXTERNAL_DOWNLOADS]: The "jarvis find" command enables searching for third-party skills on GitHub. It uses standard environment tools for web search and fetching to discover and preview "SKILL.md" files. It provides recommendations for installation but does not perform automated execution of remote code.
[DATA_EXFILTRATION]: The skill includes a "security-watch" mechanism that scans project files and git history for sensitive data such as API keys, credentials, and private keys using regular expression patterns. Findings are intended to be surfaced to the user locally as part of a security audit feature.
[PROMPT_INJECTION]: The skill implements features like "memory-recall" and "task-routing" that analyze the user's prompt and project documentation (stored in "wiki/" and ".jarvis/memory.md") to inject context-aware advice into the agent's session. This creates an indirect prompt injection surface where data from project files could influence the agent's behavior.
Ingestion points: Project wiki files ("wiki/Systems/*.md"), project memory (".jarvis/memory.md"), and general source code.
Boundary markers: The skill uses emoji-prefixed headers (e.g., "💠 JARVIS:") to label its injected context but does not implement strict delimiters to isolate untrusted data.
Capability inventory: The skill possesses extensive capabilities including full shell execution via installed hooks and the ability to modify project configuration files like "settings.json".
Sanitization: No explicit sanitization or filtering of the project data is performed before it is interpolated into the agent's prompt.

jarvis-starter