What problem does it solve?
Writing evaluation environments for Letta Code mod learning by hand is error-prone: envs must be strict JSON, distinguish real behavior from a no-op mod, and include negative controls. This Skill guides the design, drafting, and validation of mod learning env JSON files consumed by /mods learn --env.
Core Features & Use Cases
- Env Authoring Guidance: Walks through defining the target mod behavior, choosing sentinel markers, and structuring
evaluation.scenarios with happy-path, discrimination, and negative-control cases.
- Template and Validation: Ships a starter template (
assets/mod-learning-env.template.json) and a Bun validation script (scripts/validate-mod-env.ts) that checks required fields, marker arrays, memory fixtures, and scenario structure.
- Use Case: You want to teach a local mod that registers a new tool. Use this Skill to draft an env with a happy-path scenario requiring a unique sentinel in the final answer, a negative control forbidding the tool call, and forbidden trace markers catching mod load failures, then validate it before running
/mods learn.
Quick Start
Ask the agent to create and validate a mod learning env JSON for the behavior you want to teach, then run it with /mods learn --env=<path>.
