skills

Skills as a concept surface in the payload as a harness primitive for progressive disclosure, a defence against context rot, and the evidence across the source notes pulls in two directions.

Vivek Trivedy of LangChain places the idea inside a definition: agent equals model plus harness, where the harness is every prompt, tool, and hook that is not the model. Context rot is his named failure mode: performance degrades as the window fills. Progressive disclosure through skills sits alongside compaction and offloading tool results to the filesystem as one of the responses. The format is markdown, persistent, testable, shareable.

Jude Gao of Vercel tested that promise against Next.js 16 APIs missing from training data. The agent never invoked the available skill in 56% of evaluation cases, scoring the same as the no-documentation baseline. Forcing invocation with explicit instructions raised the success rate only to 79%, with results swinging on exact wording. A static «AGENTS.md» injected each turn reached 100%. Gao's reading: passive context beats active retrieval for general framework knowledge, while skills keep their use for explicit action workflows.

The other notes treat that verdict as unsettled. Jesse Vincent's «Superpowers» builds skills that are pressure-tested and extracted from technical books. Meng To counts 63 in New Form, copyable prompts he calls the ingredients to «design.md»'s recipe. Lushbinary's «Loop Engineering» guide lists skills as one of five bricks capturing project knowledge, cutting the cost of re-explaining a project.