The Agent Skills Directory

[METADATA_POISONING]: The skill makes deceptive claims regarding the source of its instructions. Specifically, reference/gpt5_prompting_guide.md claims to summarize an 'Official GPT-5 Prompting Guide' from OpenAI and provides a link to a non-existent URL (https://cookbook.openai.com/examples/gpt-5/gpt-5_prompting_guide). This misleads the agent and user about the validity and official nature of the prompting strategies provided.
[PROMPT_INJECTION]: The skill incorporates instructions that encourage the agent to override standard operational constraints. The 'Persistence Pattern' in the GPT-5 reference file directs the agent to 'Never stop or hand back when you encounter uncertainty' and 'Do not ask for confirmation'. These directives promote excessive autonomy and the bypassing of human-in-the-loop safety protocols.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process and review user-provided prompts, creating an attack surface for indirect injection.
Ingestion points: The skill accepts user prompts and requests as primary inputs in the SKILL.md workflow.
Boundary markers: There are no defined delimiters or instructions to ignore potential injections when analyzing user-provided content.
Capability inventory: The agent has the ability to read internal reference files and write output to new markdown files.
Sanitization: The instructions lack validation or escaping mechanisms for the external data being processed.

prompt-engineer