The Agent Skills Directory

[DATA_EXFILTRATION]: The skill implements a strict privacy-first architecture. All data capture is performed by a local CLI tool (scripts/learning.mjs) that writes to a .gitignore protected directory (.styleseed/learning/). The CLI explicitly forbids network transmission and raw material collection in its configuration. Furthermore, scripts/privacy-scan.mjs implements an automated guardrail to detect and reject secrets (API keys, tokens), PII (emails, IPs, card numbers), and project code/markup before any candidate record is saved.
[COMMAND_EXECUTION]: The skill uses Node.js scripts to manage design candidates. All shell execution is restricted to local management tasks (init, capture, review, prepare-share). The script scripts/secure-fs.mjs implements advanced security measures including non-blocking file opens with O_NOFOLLOW to prevent symlink attacks, directory fsync for durability, and strict basename validation to prevent path traversal.
[CREDENTIALS_UNSAFE]: While the privacy-scan.mjs script contains regex patterns for detecting secrets (like sk- or ghp_), these are used for defensive filtering (blocking the user from accidentally including secrets in design lessons) rather than storing them. No hardcoded credentials were found in the skill's own code.
[PROMPT_INJECTION]: The skill instructions in SKILL.md emphasize that agents must wait for explicit human approval before any action. It specifically notes that lessons must be human-approved and that the agent should never accept or export a candidate on a user's behalf.

ss-learn