The Agent Skills Directory

[PROMPT_INJECTION]: The skill employs sophisticated persuasion techniques designed to override agent autonomy and bypass safety or ethical filters.
Evidence in persuasion-principles.md references research titled 'Call Me A Jerk: Persuading AI to Comply with Objectionable Requests' and advocates for techniques like 'Authority' ('YOU MUST', 'No exceptions'), 'Commitment' (forcing explicit choices), and 'Scarcity' to increase compliance rates.
SKILL.md and testing-skills-with-subagents.md use high-pressure imperatives like 'Delete means delete' and 'Violating the letter of the rules is violating the spirit of the rules' to enforce behavioral discipline and prevent agents from rationalizing alternative actions.
[DYNAMIC_EXECUTION]: The script render-graphs.js performs dynamic execution by running external system binaries based on the content of other files.
Evidence: The script reads SKILL.md, extracts blocks marked as dot code, and passes them to the execSync('dot -Tsvg', ...) function. This allows the execution of the Graphviz dot utility using input dynamically retrieved from markdown files.
[COMMAND_EXECUTION]: The skill includes functionality to execute shell commands at runtime.
Evidence: render-graphs.js uses execSync to check for the existence of the dot command (which dot) and to perform graph rendering.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a surface for indirect injection where processed data could influence agent actions or script execution.
Ingestion points: The render-graphs.js script ingests content from SKILL.md (via extractDotBlocks).
Boundary markers: The script does not use specific boundary markers or 'ignore' instructions when processing dot blocks.
Capability inventory: The skill has the capability to write files and execute shell commands (execSync).
Sanitization: There is no sanitization of the content extracted from SKILL.md before it is passed to the shell-executed dot command.

writing-skills