brainstorm

The brainstorm skill acts as an intelligent research partner designed to transform ambiguous objectives into concrete technical plans for LLM reinforcement learning. Whether you are building custom environment harnesses, designing novel evaluation rubrics, or optimizing prompts via GEPA, this skill provides a structured framework to map your goals to the Verifiers ecosystem. It is intended for researchers and ML engineers looking to move rapidly from concept to experimentation by leveraging local source grounding, Prime CLI primitives, and established RL trainer workflows.

• Facilitates iterative ideation sessions that bridge the gap between high-level research intent and low-level implementation logic.

• Provides deep grounding in Verifiers architecture, including RLMEnv, BrowserEnv, and taskset design, ensuring plans are natively compatible with prime-rl.

• Nudges users toward optimal model-family selections, such as instruct-first vs. reasoning-first endpoints, and suggests repeatable experiment configurations via endpoints.toml.

• Automates the initial discovery workflow by identifying key levers: environment migration, benchmark design, prompt optimization, and RL training strategies.

• Incorporates a dedicated concept teaching mode to explain complex RL metrics, rollout tracking, and binary-reward/continuous-reward training implications.

• Start by clarifying your specific research goals, budget, and desired timeline to receive tailored environment or evaluation roadmaps.

• Use this skill to explore literature or benchmarking strategies from mid-2025 onwards, ensuring your approach remains state-of-the-art.

• Expect outputs that include structured problem framing, prioritized experiment milestones, and clear go/no-go gates for your research program.

• Before execution, ensure your workspace is prepared using prime lab setup to maintain compatibility with required CLI and python environment configurations.

• Note that this skill may suggest cloning repositories like prime-rl or prime-cli to /tmp to ensure grounding in the latest source code before drafting implementation plans.

• Always flag platform constraints and rely on first-party tooling where available to maintain project stability and ease of integration with the Prime Intellect platform.