The Agent Skills Directory

[SAFE]: The skill provides structured instructions for diagnosing common engineering failures in machine learning projects (NaN gradients, GPU OOM, slow data loading, etc.). All recommended actions use standard diagnostic tools and library functions (e.g., nvidia-smi, torch.profiler, torch.autograd.set_detect_anomaly).
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted data sources including training logs, error tracebacks, and configuration files.
Ingestion points: Reads from logs/, wandb/, runs/, and system error messages.
Boundary markers: No explicit instruction delimiters or sanitization steps are defined for processing log content.
Capability inventory: The skill has access to Bash, Write, and Edit tools.
Sanitization: No sanitization of ingested log data is specified.
Assessment: While this creates a theoretical surface for indirect prompt injection (e.g., an attacker-controlled log file containing instructions), the risk is inherent to the debugging use case and the instructions are focused on technical diagnostics rather than autonomous decision-making that could be easily subverted.

experiment-debugger