The Agent Skills Directory

[UNVERIFIABLE_DEPENDENCIES]: The skill depends on an external CLI tool that must be installed by the user via a package manager.
Evidence: SKILL.md instructs the user to execute npm install -g browse to enable the skill's core functionality.
[DYNAMIC_EXECUTION]: The skill invokes local JavaScript files using Node.js to automate data processing and report generation.
Evidence: The workflow utilizes extract_page.mjs for content scraping, list_urls.mjs for deduplication, and compile_report.mjs for the final synthesis. These are called via shell commands within the Bash Agent.
[INDIRECT_PROMPT_INJECTION]: The skill processes content from the public web to perform company research, which inherently creates a surface for indirect prompt injection if external sites contain malicious instructions.
Ingestion points: extract_page.mjs retrieves HTML content from arbitrary URLs to extract company descriptions and features.
Boundary markers: The research subagent prompt includes a dedicated "ANTI-HALLUCINATION RULES" section that explicitly instructs the model to avoid pattern-matching or letting external context override its instructions.
Capability inventory: The skill is restricted to the Bash Agent and uses it to perform file writes and execute the browse CLI.
Sanitization: extract_page.mjs contains logic to strip script, style, and noscript tags, as well as HTML comments, before presenting the text to the model.
[COMMAND_EXECUTION]: The skill makes extensive use of shell commands to manage temporary files, create output directories on the Desktop, and orchestrate the research pipeline.
Evidence: SKILL.md and references/workflow.md detail the use of mkdir, rm, cat with heredocs, and the browse tool chain.

company-research