The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes automated verification commands as part of its 'Check' workflow to validate milestones.
Evidence: The Check section in SKILL.md instructs the agent to "Show each command; run only after the user approves that exact string."
The reliance on manual user approval for the literal command string acts as a significant security control against unauthorized execution.
[DYNAMIC_EXECUTION]: The skill generates milestone commands dynamically based on previous design phases and executes them during the verification step.
Evidence: The 'Vertical Slices' phase in SKILL.md and the automated verification template in REFERENCE.md demonstrate the generation of commands from symbolic design goals.
This follows a pattern of generating executable content from templates derived from user-influenced design descriptions.
[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted user requests and existing design files to generate executable verification steps.
Ingestion points: The initial request string and existing design markdown files located at <anchor>/designs/ (specified in SKILL.md).
Boundary markers: The skill uses structured templates from REFERENCE.md to wrap content, though it lacks specific instructions to ignore embedded instructions in the ingested data.
Capability inventory: The skill can write files to the local repository (designs and ADRs) and execute shell commands (in the Check phase).
Sanitization: The skill relies on the human-in-the-loop approval process for generated commands rather than automated input sanitization.

upfront-design