Skill engineering
What it is
Skill engineering is the emerging discipline of designing, packaging, and evaluating the reusable procedures — "skills" — that AI coding agents load to do a job well. Rather than treating a skill as a one-off prompt dropped into a SKILL.md file, it treats each skill as a versioned, testable unit: a procedure layer (the instructions and examples) paired with an evaluation suite that can prove a change to the skill actually helped, run with-skill vs. without-skill to measure the delta.
Why it's emerging now
Agent Skills — the SKILL.md convention Anthropic shipped for Claude in October 2025 — spread quickly, with dozens of agent platforms adopting the format within months. But teams kept hitting the same failure mode: a skill that worked once didn't generalize, and nobody could tell whether editing a skill made the agent better or worse without re-running it against real tasks. That pushed practitioners from "just write a skill" toward treating skill design as its own discipline — specify the procedure, write evals for it, gate changes on measured pass-rate deltas — echoing the same arc that turned ad hoc prompting into "prompt engineering" and then "context engineering."
How it relates to what you know
It sits directly on top of Agent skills, already on the ledger — skill engineering is the practice built around that feature, not a new primitive. It's a sibling discipline to Harness engineering and Eval engineering: harness engineering designs the system the model runs inside, eval engineering versions the gates that judge agent behavior, and skill engineering versions the specific procedures the agent calls on — often validated with the same kind of with-skill/without-skill evaluation that eval engineering formalized.