The Agent Skills Directory

[SAFE]: The skill implements a robust authoring process that uses isolated subagent contexts for drafting and review. By ensuring reviewers have no memory of the author's reasoning, it prevents logical bias. Furthermore, the skill enforces a human-in-the-loop requirement for resolving subjective review findings, which prevents the AI from making unverified changes to project files.
[INDIRECT_PROMPT_INJECTION]: The skill processes project-related documentation files (e.g., discovery.md, proposal.md, specs/**/*.md) as grounding inputs for the AI agents.
Ingestion points: Files specified in the grounding_inputs parameter are read by the drafting and review subagents.
Boundary markers: The instructions specify that subagents are spawned in fresh isolated contexts, though they do not define specific string delimiters for the ingested content within the prompts.
Capability inventory: The skill possesses the capability to write files (during the finalize step) and spawn other internal skills (hrt-artifact-lint, humanizer, hrt-dark-mode-routing).
Sanitization: No explicit sanitization or filtering of the content within the grounding files is performed before processing.
Context: The risk associated with this injection surface is low, as the skill's primary purpose is editorial review and all non-deterministic findings must be confirmed by a human operator before being applied.

hrt-adversarial-authoring