The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses authoritative and restrictive language to mandate a specific internal workflow, effectively attempting to override the agent's standard operational logic.
Evidence: The skill defines "The Iron Law" and includes instructions such as "You MUST complete each phase before proceeding" and "Violating the letter of this process is violating the spirit of debugging."
Evidence: The provided pressure tests (test-pressure-1.md through test-pressure-3.md) explicitly challenge the agent to prioritize the skill's internal rules over conflicting external signals like time pressure, social pressure, or authority figures.
[COMMAND_EXECUTION]: The skill provides an executable utility script intended for use in local development environments.
Evidence: find-polluter.sh is a bash script that iterates through files matching a pattern and executes npm test on each to identify side effects (pollution).
[INDIRECT_PROMPT_INJECTION]: The skill encourages the deep ingestion and analysis of external, untrusted data generated during the debugging process.
Ingestion points: Phase 1, Step 1 in SKILL.md instructs the agent to "Read Error Messages Carefully," noting that "They often contain the exact solution." defense-in-depth.md and root-cause-tracing.md also suggest logging and analyzing environment variables and data flows.
Boundary markers: None. There are no instructions to use delimiters or treat the content of these external logs as potentially untrusted data.
Capability inventory: The skill provides capabilities for command execution (via find-polluter.sh) and instructions for modifying project code (via defense-in-depth.md).
Sanitization: The skill does not provide any guidance on sanitizing or escaping instructions that might be embedded within error messages, stack traces, or log outputs.

systematic-debugging