The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user input and external source text to generate requirements artifacts, creating a potential surface for indirect prompt injection. However, it implements extensive instructions to mitigate this by isolating such data within an inert-data boundary and stripping instruction-like text from summaries.
Ingestion points: User requests, pasted source code, and external documentation processed in SKILL.md and references/drafting-workflow.md.
Boundary markers: Detailed provenance labeling (outside-authored, unclear, direct-current-user) and mandatory inert-data containment for exact content.
Capability inventory: File-writing tools for specifications and records; shell commands for compliance verification as described in references/durable-records.md.
Sanitization: Normalization of untrusted content into declarative product facts and the use of escape-safe Markdown delimiters for embedded payloads.
[PERSISTENCE]: The skill includes defensive instructions in SKILL.md and references/drafting-workflow.md that explicitly forbid the agent from inspecting or modifying shell configuration files such as .bashrc, .zshrc, or .profile for persisting environment settings.
[PRIVILEGE_ESCALATION]: The agent is constrained by strict 'Write Boundaries' and a 'Read-Only-Phase Write Gate' defined in SKILL.md, which prevent unauthorized modifications to the repository, runtime state, or non-spec files.
[COMMAND_EXECUTION]: The skill allows for the execution of commands to verify compliance with decision records (references/durable-records.md), though these actions are scoped by the skill's safety gates.

vibe-requirements-spec