The Agent Skills Directory

[PROMPT_INJECTION]: The skill defines a workflow that ingests untrusted external data, creating a surface for indirect prompt injection.
Ingestion points: The agent is instructed to ingest and process external artifacts such as HAR files, log dumps, and network traces (SKILL.md), as well as interactive input from a human-in-the-loop script (scripts/hitl-loop.template.sh).
Boundary markers: There are no explicit boundary markers or instructions provided to the agent to disregard instructions potentially embedded within these ingested files.
Capability inventory: The skill utilizes broad system capabilities, including file writing (for test harnesses), network operations (via curl), and arbitrary command execution.
Sanitization: No sanitization or validation logic is specified for the external data ingested during the debugging process.
[COMMAND_EXECUTION]: The skill relies on the agent's ability to execute a wide range of commands to establish reproduction loops, including curl, git bisect, and browser automation tools like Playwright or Puppeteer.
[COMMAND_EXECUTION]: The skill involves dynamic code generation. The agent is instructed to create and run "throwaway harnesses," "minimal slices of the system," "fuzz loops," and "regression tests" to identify and fix bugs. While these are standard debugging practices, they involve the runtime generation and execution of code.

diagnosing-bugs