The Agent Skills Directory

[OBFUSCATION]: The skill includes scripts/dev-workflow.cjs, which is a large (70KB+) minified and bundled Node.js script. This distribution method obscures the underlying logic and its various third-party dependencies (such as js-yaml, gray-matter, and micromatch), making a thorough security review and verification of the implementation difficult.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a significant attack surface for indirect prompt injection due to its core research functionality.
Ingestion points: The prd-researcher sub-agent is instructed to ingest data from the local codebase and external sources via the WebFetch and WebSearch tools to identify patterns and dependencies.
Boundary markers: The skill does not define specific boundary markers or "ignore instructions" delimiters when synthesizing these research findings into the final PRD files.
Capability inventory: The skill can write files to the filesystem and execute shell/Node.js scripts. If an attacker-controlled file in the codebase or a malicious web page contains instructions disguised as "research findings," the agent may inadvertently follow them during the PRD generation or subsequent resume phases.
Sanitization: There is no evidence of sanitization or filtering of the content retrieved from external or codebase sources before it is interpolated into the prompt for synthesis.
[DATA_EXFILTRATION]: The skill automates the collection of project-specific metadata and documentation into a central directory ($HOME/.dev-wiki). It creates symlinks between project directories and this global location. While designed for productivity (a cross-project wiki), the automated aggregation of sensitive project structures and potential data leakage from multiple projects into a single home-directory location represents a data exposure risk.
[EXTERNAL_DOWNLOADS]: The prd-researcher sub-agent is configured to use WebFetch and WebSearch to retrieve information from the internet. While these are common research tools, they involve downloading content from non-whitelisted domains into the agent's context.

dev-plan