The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes a verbatim 'operating prompt' in references/operating_prompt.md that instructs the agent to bypass standard behavioral constraints. It uses phrases like 'Do not provide disclaimers to your answers' and 'Do not inform me about morals and ethics unless I specifically ask,' which are patterns designed to override the model's default safety and instructional guardrails.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user input via an interrogation workflow and a fillable worksheet.
Ingestion points: User answers to the six forcing questions in SKILL.md and assets/forcing_question_worksheet.md.
Boundary markers: The skill lacks explicit delimiters or instructions for the agent to ignore potentially malicious content embedded within these user answers during the verdict generation phase.
Capability inventory: The skill executes local command-line tools (market_first_evaluator.py, pmf_signal_scorer.py) and performs local file system writes in scripts/anti_todo_card.py to store state in ~/.andreessen-cards/.
Sanitization: No validation or sanitization of user-provided answers is implemented before they are interpolated into the prompt context.
[COMMAND_EXECUTION]: The workflow relies on the execution of multiple local Python scripts to produce deterministic scores and manage data. While these scripts use the standard library and perform arithmetic or local file storage, they represent an execution surface for the skill's logic.

andreessen