The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its core function of analyzing and extracting patterns from untrusted source code.
Ingestion points: The skill reads all project source files matching specific layer patterns in Phase 2 (P2), existing rule files in Phase 7 (P7), and the project's CLAUDE.md in Phase 7.5.
Boundary markers: The prompts for subagents (Explore, Plan, Writer agents) do not include explicit instructions to ignore embedded commands or use unique delimiters to separate instructions from the analyzed code content.
Capability inventory: The skill and its subagents have access to Read, Write, Edit, Bash, Agent, and Skill tools, which could be abused if an injection is successful.
Sanitization: No sanitization, escaping, or filtering of the extracted content is performed before it is processed or written to convention documents and rules.
[COMMAND_EXECUTION]: The skill executes local shell scripts to perform project discovery, stack detection, and environment setup.
Evidence: Execution of scripts/convention.sh with subcommands such as detect-stack, scan, setup, and validate via the Bash tool.
Context: These commands are used for legitimate project scanning and setup tasks within the plugin's own directory structure.
[SAFE]: The skill utilizes vendor-specific subagents and documentation resources authored by the vendor.
Evidence: Integration with brewtools:text-optimizer, brewcode:bc-rules-organizer, and references to doc-claude.brewcode.app.
Context: These are internal components of the brewcode ecosystem by kochetkov-ma and represent expected vendor functionality.

convention