The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructions involve the ingestion and analysis of external data, including model checkpoints, training datasets, and experimental result records.
Ingestion points: Dataset splits, model outputs, and result records are processed in SKILL.md and references/training-and-recovery.md.
Boundary markers: None explicitly defined for untrusted data analysis.
Capability inventory: The skill guides the agent in dispatching evaluation campaigns and monitoring training runs.
Sanitization: The skill assumes data is part of a research environment but provides no specific sanitization instructions for model outputs.
[EXTERNAL_DOWNLOADS]: The skill references established research materials from trusted sources.
Evidence: references/sources.md references the Google Deep Learning Tuning Playbook at github.com/google-research/tuning_playbook.
Evidence: references/sources.md references a research paper hosted on arxiv.org.
[NO_CODE]: The skill consists entirely of markdown documentation and instructional guidelines. It contains no scripts, executables, or dynamic configuration files.
[SAFE]: The skill explicitly promotes secure practices. Specifically, references/training-and-recovery.md instructs the agent to 'Store credentials in their configured mechanism, not experiment records', which prevents accidental credential exposure in logs.

credo-experiment