The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands (open, xdg-open, or start) to automatically open generated HTML reports. The file paths for these reports are constructed using user-provided company names (offer-compare-<companyA>-vs-<companyB>-<YYYYMM>.html). This pattern creates a potential command injection vulnerability if the company names contain shell metacharacters (e.g., ;, &, |) and the agent platform does not strictly sanitize the input before shell execution.
[DATA_EXFILTRATION]: While the skill primarily operates locally, it encourages the aggregation of highly sensitive personal and financial data, including specific salary breakdowns, equity vest schedules, visa status, and family situations. It stores this data in plain text within the offer-bank/ directory inside the skill folder and in a public user folder (~/Desktop/Claude skills/). This practice increases the risk of sensitive data exposure to other local processes or users.
[INDIRECT_PROMPT_INJECTION]: The skill ingests raw, untrusted data from multiple job offers and interpolates it directly into various analysis prompts and the final HTML report. There are no explicit instructions for the agent to sanitize or escape this input. A malicious payload within a job offer description could potentially influence the agent's logic during the breakdown or risk analysis phases, or inject malicious scripts into the generated HTML file that is automatically opened in the user's browser.
[PROMPT_INJECTION]: The skill contains 'Iron Rules' that explicitly command the agent to override standard neutral behavior. It mandates that the agent 'must give a clear recommendation' and forbids neutral responses like 'both look great,' which forces the AI to abandon its objective analytical guardrails.

offer-compare-skill