The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a 'Research -> Plan -> Do' pipeline where external information gathered from the web is used to generate subsequent instructions and code.
Ingestion points: External data from web searches or documentation fetches is stored in files such as FINDINGS.md and .prompts/*/auth-research.md, which are then used as context for planning and implementation.
Boundary markers: While templates use XML tags like <research> and <findings>, there are no explicit instructions or 'ignore-previous-instructions' delimiters to prevent instructions embedded in external content from hijacking the agent.
Capability inventory: The generated prompts have the capability to write/modify files and execute shell commands via the agent's standard toolset.
Sanitization: The skill does not specify any sanitization or validation logic for external content before it is interpolated into the executable 'Plan' prompts.
[DYNAMIC_EXECUTION]: The skill is designed to dynamically generate PLAN.md files which it defines as 'executable prompts'. These generated documents are not merely records but are intended to be loaded and executed as the primary instruction set for subagents or subsequent sessions.
[DYNAMIC_CONTEXT_INJECTION]: The reference/meta-prompts.md file defines a research template that uses the shell expansion syntax !date +%Y-%m-%d`` to inject the current date into the agent's context at runtime. This allows for silent shell command execution during the prompt initialization phase.
[COMMAND_EXECUTION]: The skill utilizes basic shell commands (mkdir, ls, find, mv) to manage its hierarchical directory structure and archive completed prompts. It also writes execution metrics to ~/.claude/skill-analytics/ upon completion.

planning-prompts