The Agent Skills Directory

[SAFE]: The skill implements a robust, multi-stage process for authoring slash commands with built-in safety devices (verifiers and hardstops) to prevent common AI failure modes.
[SAFE]: All internal logic is executed via local Python scripts (tools/classify_signals.py, tools/check_safety.py) and static markdown reference files. No external resources are downloaded or executed at runtime.
[SAFE]: The tool employs a deterministic safety gate (tools/check_safety.py) using jsonschema to enforce that every generated command includes non-empty verifier and hardstop definitions.
[SAFE]: The subprocess.run call identified by static analysis in tools/test_check_safety.py is a test utility that runs the skill's own validation script and does not pose a security risk as it does not directly process untrusted user input in the shell.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface as it ingests user-provided task descriptions and interpolates them into prompt templates to generate final slash commands. This is an inherent part of its function as an authoring tool.
Ingestion points: User input for task details and evaluation criteria collected in SKILL.md and references/interviewer.md.
Boundary markers: The generated commands are organized into structured markdown sections (## Procedure, ## Verifier (auto)) as defined in references/renderers/claude-code.md.
Capability inventory: The skill has the capability to write markdown files to the local .claude/commands directory.
Sanitization: The skill performs slugging for filenames but otherwise performs direct substitution of user text into template skeletons.

loop-forge