brainstorm

The brainstorm skill is a structured, interactive research companion designed for AI engineers and researchers working within the Prime Intellect ecosystem. It serves as an ideation engine that bridges the gap between abstract research objectives and the technical implementation of RL training, benchmark construction, and agent environment design. By grounding its suggestions in local repository sources and industry-standard workflows, it ensures that proposed solutions are highly actionable and compatible with the user's specific codebase, whether it involves simple scripted tasks or complex, vision-based browser automation using Computer Use Agent (CUA) or DOM-based primitives. The skill is engineered to facilitate deep technical exploration, making it an essential tool for those building robust evaluation harnesses, optimizing prompt strategies via GEPA (Goal-Oriented Prompt Augmentation), or refining RL reward functions. It actively manages the research lifecycle by recommending appropriate model families for different stages of exploration—from instruct-first prototyping to reasoning-first deep dives—while ensuring alignment with the project's budget and technical constraints.

• Facilitates iterative collaborative research cycles instead of static one-shot planning.

• Maps high-level research goals to technical levers such as environment migration, benchmark design, and reward function refinement.

• Provides context-aware guidance by analyzing local CLI configurations, verifiers workspace files, and Prime-RL components.

• Recommends model-family-specific strategies, including guidance on leveraging gpt-4.1, qwen3, and reasoning-focused models like glm.

• Supports advanced RL development, including truncation, branching trajectories, and binary-reward vs. continuous-reward trade-offs.

• Offers structured output including problem framing, value-ranked experiment plans, milestone definition, and go/no-go gate analysis.

• Always clarify the model family, time horizon, and specific research objective before beginning deep planning.

• Use the brainstorm skill to scan for recent academic benchmarks and papers, prioritizing mid-2025 and later literature unless historical context is requested.

• Leverage the tool for guidance on deploying environments to the Environments Hub, including instructions for public vs. private visibility toggles.

• When blocked by platform limitations, the agent will pause to request user intervention rather than proceeding with hidden assumptions about scoring contracts or prompt formatting.

• Always anchor technical explanations in native terminology, such as trajectory-based tracking or RLMEnv context management, to ensure consistent, grounded communication with the Prime Intellect infrastructure.