The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data such as session logs, transcripts, and third-party skill files (e.g., Cursor rules, GPT instructions) to generate or convert BMad skills.
Ingestion points: Data is ingested through the 'from-logs' approach (help/approaches.md) and the 'convert' mode (help/how-to-convert-external.md), which reads external files and pasted text.
Boundary markers: The documentation describes a 'discovery conversation' and a 'read-back' process for human approval but does not specify technical delimiters to prevent the agent from executing instructions embedded in the source data.
Capability inventory: The system can write files to the filesystem (help/how-to-personal-skill.md), execute Python scripts via 'uv run' (help/craft-scripts.md), and perform automated evaluations that bypass standard CLI permission prompts (help/evals.md).
Sanitization: The skill includes a review mode (help/how-to-review.md) that applies lenses for architecture and leanness, though its primary focus is structural rather than malicious instruction filtering.
[COMMAND_EXECUTION]: The module documentation specifies the use of 'uv run' to execute Python scripts associated with skill workflows. It also describes an evaluation harness (bmad-eval) designed to run non-interactive agent sessions for testing purposes, which intentionally bypasses standard CLI permission prompts.
[DYNAMIC_EXECUTION]: The 'Rendered skill' shape utilizes a templating system and a specialized script (render_skill.py) to generate agent instructions at runtime by merging base skill content with user-specific and team-specific customization files.
[EXTERNAL_DOWNLOADS]: The configuration in bmod.toml specifies a GitHub repository owned by the authoring organization (bmad-code-org/BMAD-METHOD) as the source for module updates.

bmod-toolsmith