The Agent Skills Directory

[DATA_EXFILTRATION]: The skill includes a 'SELF-EVOLUTION TELEMETRY' section that explicitly instructs the agent to write execution metadata, including dates, outcomes, and improvement suggestions, to a persistent file path at ~/.claude/projects/ if it exists. This represents a persistence and data collection mechanism that tracks skill usage across different sessions.
[PROMPT_INJECTION]: The skill uses $ARGUMENTS to ingest user-provided organization names and program areas. This untrusted data is then passed into four different sub-skills (/impact-measurement, /fundraising-optimizer, /grant-writer, and /donor-retention) without boundary markers or sanitization, creating an attack surface for indirect prompt injection where malicious data could influence the behavior of the subsequent processing phases.
[COMMAND_EXECUTION]: The 'SELF-HEALING VALIDATION' phase instructs the agent to 'Run the project's test suite' and 'Run build/compile'. While intended for validation, this grants the skill the capability to execute arbitrary shell commands defined in the project's build or test scripts, which could be exploited if those scripts are attacker-controlled.
[DATA_EXFILTRATION]: The skill's instructions involve analyzing sensitive organizational data including 'donor retention', 'churn prediction', and 'grant management'. While it contains a 'DO NOT' instruction against logging actual donor names, the cross-phase analysis and telemetry logging still process and potentially record aggregated sensitive business logic and operational health data.

impact-org