Security audit
Security checks for vulnerabilities and agentic risk
This skill is a disclosed product-marketing setup helper that reads project materials to draft a local context document and does not include hidden code, network behavior, or credential handling.
Before using auto-draft, be aware it may read project files and summarize product, audience, pricing, and positioning into .agents/product-marketing-context.md. Review the draft before relying on it or sharing it, especially if the repository contains private strategy, customer names, pricing, or internal notes.
Skill instructs the agent to omit warnings, disclaimers, or ethical commentary. Stripping safety caveats hides risk from the user and is a common jailbreak preamble.
- Voice do's and don'ts (1-2 each): e.g., "We explain, we don't lecture" or "We're confident, never arrogant"
The instruction to use the skill 'at the start of any new project' is an ambiguous global activation condition that can override normal intent matching and cause the skill to run in many unrelated project starts. In context, this is risky because the skill can read repository materials and write .agents/product-marketing-context.md, so accidental invocation expands access and side effects without sufficiently specific user direction.
The instruction to use the skill 'at the start of any new project' is an ambiguous global activation condition that can override normal intent matching and cause the skill to run in many unrelated project starts. In context, this is risky because the skill can read repository materials and write .agents/product-marketing-context.md, so accidental invocation expands access and side effects without sufficiently specific user direction.
This JSON eval file is a manifest-like file, so vague-trigger checks apply. The expected behavior explicitly says the skill 'Should trigger on casual phrasing' for a generic request to 'create a product context doc', which overlaps with ordinary conversation and does not define clear activation boundaries or exclusions.
The auto-draft workflow instructs the agent to scan the codebase and marketing materials, then draft and save context, but does not require clear disclosure of what files will be accessed or explicit approval before scanning and writing. This can lead to overbroad repository access and unanticipated persistence of derived content, which is more sensitive than a simple manual draft flow.
The skill directs the agent to study the repo—README, landing pages, package.json, and other docs—to auto-draft the context, but it does not require an explicit warning or consent flow about scanning repository contents. Even though this is framed as a productivity feature, it can surprise users by broadening data access and pulling potentially sensitive internal information into generated marketing context.
The eval expects the agent to create .agents/product-marketing-context.md when complete, but does not require an explicit confirmation step immediately before writing. Silent writes can cause unintended filesystem changes, especially if the conversation produced a draft the user did not intend to save yet.
No suspicious patterns detected.