The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill reads and processes content from multiple external project files, creating a surface for indirect prompt injection where malicious instructions in design documents could influence agent behavior.
Ingestion points: The skill reads CLAUDE.md, design/gdd/game-concept.md, design/gdd/systems-index.md, docs/architecture/architecture.md, and docs/architecture/control-manifest.md in Phase 1 to define its scope and technical constraints.
Boundary markers: Absent. There are no delimiters or explicit instructions provided to the agent to treat the ingested project content as untrusted data.
Capability inventory: The skill possesses file-writing capabilities (Write, Edit) to specific project directories, shell execution (Bash) restricted to a project-local helper script, and the ability to spawn sub-agents (Agent) for creative reviews.
Sanitization: Absent. No sanitization, escaping, or validation is performed on the content of the ingested markdown files before they are interpreted.
[DYNAMIC_CONTEXT_INJECTION]: The skill utilizes dynamic context injection to resolve configuration values at the moment the skill is loaded.
Evidence: The skill starts with !bash "${CLAUDE_SKILL_DIR}/../../hooks/yaml-helper.sh" resolve_config --keys review_mode,automation,workflow, which executes a shell command silently when the file is opened.
[COMMAND_EXECUTION]: The skill relies on the execution of a local shell script residing outside of the skill's specific directory.
Evidence: The skill is configured to allow and invoke bash "*/.claude/skills/vertical-slice/../../hooks/yaml-helper.sh" resolve_config *.
Note: This script appears to be a project-level utility, but its execution means the skill is dependent on external executable code whose content is not provided within the skill itself.

vertical-slice