The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's core workflow involves ingesting training/validation data and evaluation feedback to rewrite agent instructions.
Ingestion points: Data loaders (load_trainset, load_valset) and the evaluation result interface (evaluate_example) in SKILL.md.
Boundary markers: The documentation recommends keeping evaluators separate and deterministic, but the framework lacks explicit technical delimiters to prevent adversarial instructions in the input data from influencing the proposer loop.
Capability inventory: The optimized RLM (PredictRLM) can execute tools, access sandboxes, and perform document analysis tasks.
Sanitization: The provided templates do not include sanitization logic for the examples or feedback strings that drive the instruction optimization process.
[DYNAMIC_EXECUTION]: The framework dynamically updates and executes instructions within the PredictRLM instance based on the output of an optimization proposer.
Implementation: As shown in the DocumentProject class, the evaluate_example method instantiates and executes a new RLM with candidate instructions generated at runtime. This allows for programmatic modification of the agent's logic based on external inputs.

rlm-gepa