The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses strong instructional language to enforce a strict multi-step workflow and quality gates (e.g., "CRITICAL", "ENFORCEMENT", "[STOP — BLOCKING]"). These instructions are functional and intended to ensure the agent follows the orchestration process correctly rather than attempting to bypass underlying safety filters.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user requirements provided via the $ARGUMENTS variable to bootstrap the codebase analysis phase.
Ingestion points: User input enters the workflow in SKILL.md (Step 1) and is passed to sub-agents in subsequent steps.
Boundary markers: The skill does not use explicit XML or markdown delimiters to isolate user input from instructions when extracting search keywords.
Capability inventory: The skill utilizes filesystem search tools (rg, grep) and spawns multiple sub-agents (codebase-analyzer, technical-designer, etc.) with the capability to read files and generate documentation.
Sanitization: The instruction to "Extract candidate keywords" (feature names, domain nouns, etc.) before performing searches serves as an effective manual sanitization step performed by the agent, preventing the direct passage of raw user input into tool arguments.
[COMMAND_EXECUTION]: The skill directs the agent to execute local search commands (rg or grep) to locate relevant files. These commands include explicit exclusions for sensitive or irrelevant directories like .git and node_modules, focusing the tool usage on the intended codebase analysis.
[DATA_EXPOSURE]: The workflow facilitates reading and analyzing the local codebase to generate technical design documents. This is the primary intended purpose of the skill. The process includes a mandatory "Scope Confirmation" gate where the user must explicitly approve the identified files before any design work or deeper analysis occurs.

recipe-design