The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes several shell commands during its initialization phase to search for configuration files and create the vault directory structure. These include find, grep, ls, cat, and mkdir operations.- [DATA_EXFILTRATION]: The skill performs a silent context scan that accesses files outside of the current working directory. Specifically, it reads from ~/.claude/CLAUDE.md, scans the user's home directory for other skills in ~/.claude/skills/, and traverses parent directories to find existing configuration files. While these files contain instructions rather than secrets, accessing the user's home directory configuration without explicit prompting is a data exposure risk.- [PROMPT_INJECTION]: The skill utilizes the WebFetch tool to scrape content from external websites and LinkedIn profiles provided by the user. This external data is then incorporated into 'Context' files such as voice.md and business.md. The vault's core configuration file (claude.md) is designed to load these files as system instructions at the start of every session, creating a vulnerability where malicious instructions on a website could influence the agent's long-term behavior via indirect prompt injection.
Ingestion points: Website and bio content scraped during Phase B, Step B.3 in SKILL.md.
Boundary markers: The markdown templates in references/ do not use delimiters or instructions to treat the scraped content as untrusted data.
Capability inventory: The skill possesses the capability to read/write to the filesystem, execute shell commands, and perform network requests as defined in SKILL.md and references/settings-json-template.md.
Sanitization: There is no evidence of validation or sanitization of the scraped external content before it is processed and stored as system-level context.

setup