The Agent Skills Directory

[PROMPT_INJECTION]: The skill incorporates safeguards against instruction overrides from scraped content. 'Hard Rule 2' (Read facts, never copy prose) specifically directs the agent to describe observed behaviors in its own words rather than transcribing verbatim text, which helps mitigate the risk of processing malicious instructions embedded in competitor sites or artifacts.
[DATA_EXFILTRATION]: The skill manages potentially sensitive competitor data. It emphasizes PII handling discipline (Rule 5), requiring redaction of personal information. It also provides specific guidance on securing browser session tokens (storageState) by adding them to .gitignore, and mandates maintaining the privacy of reference corpora.
[COMMAND_EXECUTION]: The skill outlines the use of browser automation tools (e.g., Playwright, Claude computer-use) for walkthroughs. These operations are conducted within an 'assume-malicious' framework, using throwaway records and requiring owner-assisted fallback for inaccessible or sensitive areas, ensuring execution remains within authorized boundaries.
[PROMPT_INJECTION]: (Indirect Prompt Injection Analysis)
Ingestion points: Untrusted data enters the agent context through browser-based walkthroughs of competitor products, mining of public documentation repositories (e.g., Mintlify, Docusaurus), and analysis of client-side JavaScript bundles.
Boundary markers: The skill uses clear instructional boundaries, repeatedly telling the agent to 'read facts, never copy prose' and to describe observed behavior in its own language.
Capability inventory: The skill utilizes browser automation tools (computer-use, Playwright) and file system interactions for report generation and screenshot organization.
Sanitization: Directives include strict PII redaction, human-in-the-loop verification for sensitive captures, and instructions to log and route any discovered secrets to security guardians.

competitor-recon-weapon