The Agent Skills Directory

[SAFE]: The skill implements a 'Preconditions' and 'Red flags' framework that emphasizes human-in-the-loop decision-making and tool-assisted verification.
[SAFE]: The use of subagents for 'Probe before you write' operations (specifically the Overlap and Architecture probes) is a strong defensive design. By ensuring the main agent context does not ingest entire requirement ledgers or system architecture documents (SAD), the skill minimizes the risk of context window exhaustion or the introduction of conflicting/malicious instructions from large local files.
[SAFE]: Requirement IDs are minted via a script (new-id.sh), and the skill explicitly forbids manual ID creation, which prevents ID collision attacks or unauthorized metadata modification.
[SAFE]: The 'Supersession' logic described in references/supersession.md enforces a strict reciprocal relationship between old and new requirements, preventing 'dangling' or 'silent' requirement changes that could bypass safety gates.
[SAFE]: The glossary management rules in docs/CONTEXT.md promote clear boundaries and prevent ambiguity-based prompt injection attacks where a user might try to redefine safety-critical terms to mean something benign.
[SAFE]: Network or file operations (like check-units.sh or find-items.sh) are scoped to local repository maintenance and do not indicate data exfiltration or unauthorized remote access.

grill-requirements