idea-create

Define your compute, time, and model-size constraints first, then search across foundation model, application, and interdisciplinary spaces rather than only LLM-centric topics. This Skill produces 8-10 candidates with minimal viable experiments and converges on 3-5 fully developed proposals.

Feasible directions include inference-time methods, evaluation and mechanism studies, failure mode discovery, and lightweight fine-tuning with LoRA on 7B-32B open models. The Skill explicitly excludes pretraining, heavy RL post-training, and large-scale benchmark construction.

Yes, it mandates coverage of AI plus other disciplines such as AI4Science, healthcare, law, finance, education, and social science. A final coverage audit checks whether interdisciplinary candidates were overlooked and adds supplements if needed.

Each top candidate must identify a core theoretical pillar, such as definitions, propositions, error decompositions, or mechanistic explanations, alongside experiments that directly validate the main claim. Ideas offering only decorative theory or table-stacking experiments are downgraded or excluded.

It does not execute experiments or access literature databases, so novelty must be verified against recent publications. It also excludes directions needing large-scale training, heavy infrastructure, or more than roughly three days to reach a closed-loop draft.