The Agent Skills Directory

[SAFE]: The skill implements robust security guardrails for external actions. Instruction Step 10 explicitly mandates that for external writes, publications, messages, or payments, the agent must resolve the exact target, scope, and authorization first, specifically stating that a 'send' instruction is not authorization for an unresolved recipient.
[SAFE]: The skill employs a transparent 'reification record' system that logs every signal, source, decision, and assumption. This record provides full visibility into the agent's reasoning and state, including a structured decision ID system (D-001) that supports reversibility and prevents hidden choices.
[SAFE]: Script execution is limited to local validation tools. The skill uses a Python script (validate_brief.py) to ensure the integrity of structured handoff data (BRIEF.json). The static analysis finding of subprocess.run in the test suite is a benign use of list-based command execution for testing local utilities and does not expose the system to shell injection.
[SAFE]: The instructions contain strict prohibitions against 'hallucinating' or inventing biographical facts. Step 7 and the 'thin-facts-no-invention' example ensure that every personal detail in a draft must trace back to user-supplied text or cited sources, using bracketed slots for missing information rather than inventing data.
[SAFE]: The skill adheres to the 'least privilege' principle by discovery of capabilities only when needed and reading contracts before usage (Step 5), separating idea shaping from actual side effects.

reify