The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructs the agent to ingest untrusted external data, creating an attack surface for indirect prompt injection.
Ingestion points: In stages/01-investigation/investigation-guide.md, the agent is directed to read and analyze external source code, documentation, papers, and dataset manifests.
Boundary markers: The workflow templates (e.g., stages/01-investigation/investigation-notes-template.md) lack explicit delimiters or instructions to ignore or isolate potential prompts embedded within the ingested external content.
Capability inventory: The skill utilizes capabilities including file system modification (Stage 4) and command execution for training and evaluation (Stage 5), as documented in stages/04-implementation/implementation-template.md and stages/05-training-and-validation/training-validation-template.md.
Sanitization: There is no requirement or instruction for the agent to sanitize or validate retrieving content from external sources before it influences code generation or execution.
[COMMAND_EXECUTION]: The skill facilitates the generation and subsequent execution of shell commands and scripts.
Evidence: stages/04-implementation/implementation-template.md tracks the creation and modification of scripts and configs in the Change Inventory, while stages/05-training-and-validation/training-validation-template.md records the exact commands executed for training and validation runs.
[NO_CODE]: The skill itself consists of Markdown guides and YAML configuration without containing executable scripts or binary files, though it instructs the agent to produce and run them.

llm-fine-tuning-skill