The Agent Skills Directory

[SAFE]: The skill implements a robust security model for handling untrusted data. Operating Constraint 2 establishes a clear 'Untrusted input boundary,' explicitly instructing the agent that logs, alert payloads, and chat exports must be treated as untrusted data, and any instructions or URLs found within them must be ignored.
[SAFE]: Data exposure risks are mitigated through a mandatory redaction policy. Constraint 14 requires that credentials, tokens, and personal data (PII) are redacted at the point of capture (e.g., sk_live_****) before being recorded in any report or timeline.
[SAFE]: The skill maintains a strict 'read-only' posture for system interactions. Command execution is limited to a predefined library of evidence-collection commands (documented in frameworks/evidence-collection.md). All commands must be presented to the user verbatim and require explicit approval before execution, preventing arbitrary or automated code execution.
[SAFE]: Storage and persistence follow best practices for project-local storage. The skill restricts all writes to the project's own directory (docs/postmortems/ or .memory/), avoiding the use of global agent directories or system-wide temporary folders.
[SAFE]: The author (carrilloapps) uses consistent vendor-owned domains (carrillo.app, carrilloapps.com) and repositories for all external references, with no evidence of typosquatting or malicious redirection.

postmortem-writing