The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses forceful instructional language such as 'MANDATORY IMPORTANT MUST CRITICAL' and 'NO EXCEPTIONS' to ensure strict adherence to its internal protocols. It also contains instructions to 'Ignore Claude-specific mode-switch instructions' and a self-authorizing clause for subagents: 'that skill activation authorizes use of the required spawn_agent subagent(s) for that task'.
[COMMAND_EXECUTION]: The instructions direct the agent to perform multiple shell-based operations, including grep, rg, find, git ls-files, and ls -d, to scout the repository. It also calls for the execution of a local script, python .claude/scripts/code_graph, to trace code connections.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted data from the repository, such as source code and documentation, as well as user-provided requirements. This data is then used to generate feature specifications and test cases without explicit sanitization or instructions to ignore potentially malicious embedded content.
Ingestion points: The skill reads repository source code and documentation via grep, rg, and git commands described in the 'INIT' and 'UPDATE' modes in references/author.md.
Boundary markers: There are no explicit boundary markers or instructions to the agent to disregard instructions found within the ingested codebase files.
Capability inventory: The skill possesses the capability to read files, execute shell commands, run local Python scripts, and write to the filesystem to update documentation.
Sanitization: No specific sanitization, validation, or filtering of the ingested data is described before it is incorporated into the generated artifacts.

spec