The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied fragment files, briefs, and editorial directions through multiple subagent roles (assembler, polisher, reviewer). Malicious instructions within these user-provided documents could potentially influence the agent's behavior.
Ingestion points: Fragments file, Brief, and Editorial direction inputs, which are copied to /tmp/mosaic--/ as fragments.md, source.md, and feedback.md (specified in the 'Working Files' section of SKILL.md).
Boundary markers: The skill instructs the orchestrator to pass file paths rather than content strings to subagents, but it does not define specific XML-style delimiters or clear 'ignore' instructions for the ingested text content.
Capability inventory: The skill has capabilities to write files (Write tool), execute shell scripts (lint.sh, frequency.py), and commit changes to the repository.
Sanitization: There is no explicit requirement for the agent to sanitize or validate the content of the user fragments before processing.
[COMMAND_EXECUTION]: The orchestrator is instructed to execute local utility scripts via the shell to perform mechanical checks on the draft.
Evidence: Execution of ../catalog-deslop/references/lint.sh and python3 ../catalog-deslop/references/frequency.py on the draft.md file (specified in the 'Stage A' and 'Stage C' sections of SKILL.md).
Context: These scripts are part of the vendor's own suite of skills (av/skills) and are used for linting and frequency analysis.

mosaic-writing