The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external text data from the web for prompt evaluation purposes.
Ingestion points: Processes "verified real text" from Taiwanese websites, news, and public posts as described in references/sources.md and SKILL.md.
Boundary markers: No specific boundary markers or sanitization instructions are defined for the corpus text data within the prompt construction.
Capability inventory: The skill executes local scripts (sweep.mjs), uses playwright-cli to control a browser, and performs npm install for dependencies.
Sanitization: No explicit sanitization or escaping of external content is mentioned before it is interpolated into model sessions.
[EXTERNAL_DOWNLOADS]: The skill requires downloading external packages and references third-party repositories.
Evidence: references/setup.md instructs the user to run npm install and npm install -g @playwright/cli@latest to set up the environment.
Trusted Source: References the Playwright CLI repository at github.com/microsoft/playwright-cli, which belongs to a trusted organization.
[COMMAND_EXECUTION]: The skill relies on executing various local commands and scripts to manage experiments.
Evidence: Instructions in SKILL.md and references/setup.md include running npm run build:extension, scripts/prompt-experiments/sweep.mjs, and playwright-cli attach to interact with the local browser.
[DATA_EXFILTRATION]: The skill accesses sensitive local file paths to configure the connection to the Chrome browser.
Evidence: references/setup.md instructs the agent to read the DevToolsActivePort file located within the Chrome profile directory (e.g., ~/Library/Application Support/Google/Chrome Beta/).
Context: While this access is necessary for establishing a Chrome DevTools Protocol (CDP) connection for the tool's primary purpose, these directories contain sensitive user data.

prompt-experiments