+the largest cluster by count, covering the full spectrum from fundamental research to practical developer tooling. the unifying theme: making AI systems more capable, reliable, and useful — whether by improving the models themselves ([[llm-behavior-improvement|LLM behavior improvement]], [[flapping-airplanes|AI training efficiency]]) or by building the infrastructure around them ([[spec-driven-dev|spec-driven dev kit]], [[context-window-optimizer|context window optimizer]], [[hard-docs-writer|hard docs writer]]).

+the most actionable ideas in this cluster are the developer tooling ones: [[spec-driven-dev|spec-driven dev kit]] (rated [DO THIS] — research → plan → implement pipeline with context management) and [[overnight-app-grinder|overnight app grinder]] (autonomous coding agent manager). both reflect a meta-insight: the bottleneck for AI-assisted development is not model capability but workflow — how you structure the problem, manage context, and review outputs. [[agents-md-research|AGENTS.md optimization research]] goes even deeper, asking how instruction structure affects model recall. on the application side, [[ai-agent-reply|AI agent reply]] and [[ai-conversationalist|AI conversationalist]] both depend on [[me-model|me model]] for personalization, while [[ai-onboarding|AI onboarding]] addresses the human adoption side.

+

+the more speculative work — [[cognitive-foom|cognitive foom]] (recursive self-improvement infrastructure) and [[llm-physical-intuition|LLM physical intuition]] — is research-oriented and harder to scope into a 2-month project, but potentially more impactful. [[context-window-optimizer|context window optimizer]] is the connective tissue between this cluster and [[cluster-memory-and-context|memory and context tools]] — if you're building agents that work with personal context ([[axon|axon]], [[always-on-ai-assistant|always-on assistant]]), context management is a first-class engineering concern.