The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external artifacts such as HAR files, log dumps, and event traces which may contain embedded malicious instructions. Ingestion points: SKILL.md Phase 1 identifies logs, HAR files, and traces as inputs. Boundary markers: Lacks explicit delimiters but provides strong textual instructions to redact secrets and isolate signal lines. Capability inventory: The skill executes shell commands, generates test scripts, and performs network requests. Sanitization: Includes a dedicated 'Redact' phase requiring the replacement of secrets with placeholders before processing.
[DYNAMIC_EXECUTION]: The skill generates and executes scripts at runtime, including test harnesses, fuzzing loops, and headless browser scripts (Playwright/Puppeteer) to reproduce identified issues.
[COMMAND_EXECUTION]: The workflow utilizes several shell-based tools and commands, including curl, git bisect, and a template-based human-in-the-loop script (scripts/hitl-loop.template.sh).
[EXTERNAL_DOWNLOADS]: The skill documentation (UPSTREAM.md) references an external repository at github.com/mattpocock/skills for tracking provenance and facilitating maintenance updates.

diagnosing-bugs