skills

Vivek Trivedy of LangChain classes skills as a harness primitive for progressive disclosure, one of the defences against context rot: capability loaded when the task calls for it instead of sitting in the window every turn. The carrier is deliberately plain, persistent markdown that is testable and shareable (fichiers SKILL.md). Jesse Vincent built Superpowers on that premise in October 2025, treating a skill as a module the agent reads and improves, pressure-tested against extreme scenarios and sometimes extracted from technical books.

Jude Gao's evaluation at Vercel, published 27 January 2026, lands on the weak joint. In 56% of evaluation cases the agent never invoked the skill that was available to it, producing the same 53% success rate as the baseline with no documentation. Adding explicit instructions that forced the trigger lifted success only to 79%, and results swung with the exact wording. A static AGENTS.md injected into the system prompt every turn reached 100%. The cost of the concept sits at the decision point: something has to choose to retrieve.

The framing holds elsewhere. Meng To's formula is design.md as recipe, HTML as finished dish, skills as ingredients, 63 of them in New Form. Lushbinary counts skills among the five bricks of Loop Engineering, the layer that stops project context being re-explained. Daniel Meppiel, writing on 1 October 2026, wants them handled as software: versioned packages, owners, approval rules, a supply chain, rather than copied files. Governance of a catalogue is a different problem from invocation.