The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests data from multiple local file sources and interpolates it into sub-agent prompts.
Ingestion points: The skill reads project manifests, technical specifications, implementation plans, and user-provided guidance from .claude/conclave/guidance/ and .claude/conclave/eval-examples/ (SKILL.md, Setup steps 3-16).
Boundary markers: While it uses delimiters and headers like ## User Project Guidance (informational only), these markers do not provide absolute protection against adversarial instructions embedded in documentation.
Capability inventory: The sub-agents created by this skill have the ability to write code, modify files, and execute runtime tests via the environment, which could be misused if they follow malicious instructions found in project files.
Sanitization: Instructions explicitly mandate that guidance content be injected verbatim and not filtered or reinterpreted, preventing defensive sanitization of untrusted project data.
[COMMAND_EXECUTION]: The skill facilitates the execution of code through its sub-agents, particularly the QA agent which is tasked with writing and running Playwright tests against the application (SKILL.md, Orchestration Flow step 6). While this is the intended purpose of the engineering skill, it represents a significant capability surface that depends on the integrity of the input specifications provided in the project directory.

build-implementation