Most context-engineering advice assumes the bottleneck is how many tokens a model can hold. This collection starts from the opposite claim: the binding constraint is attention, not capacity, and long contexts fail predictably through lost-in-the-middle effects and attention scarcity well before the window is full. Its core move is to treat curation as a reusable, versioned skill rather than a one-off prompt you rewrite for every project.
What Sets It Apart
- Fifteen skills span foundations, architecture, operations, and cognitive modeling, surfaced through progressive disclosure so only the relevant one enters the window—the design eats its own dog food on context budgeting.
- A routing mechanism decides which skill fires for a given task; it was benchmarked across four frontier models with measured effect sizes, reaching 0.92 top-1 accuracy on newer ones, so selection is empirical rather than guesswork.
- A
researcher/operating system keeps append-only ledgers with novelty gates and human oversight, giving the set institutional memory and a data-driven path to evolve instead of bit-rotting. - Skills are written as Python pseudocode and target Claude Code, Cursor, Codex, and Copilot alike, so the ideas are not tied to one vendor's runtime.
Who It's For
Great fit if you build multi-step agents and keep hitting quality cliffs as context grows, or if you want a vendor-neutral vocabulary for context engineering backed by measured benchmarks. Look elsewhere if you expect a drop-in library—these are conceptual skills plus pseudocode you adapt by hand, not an installable framework, and the benchmark numbers come from the author's own harness rather than independent replication.
