Somebody else decided what good looks like, and it shipped with the skill you installed. Here’s the guide to fix it.

Source: Nate’s Newsletter Date: 2026-08-01 URL: https://natesnewsletter.substack.com/p/agent-skill-one-job-test

Summary

Nate argues that installing an agent skill imports someone else’s embedded decisions about what “good” output and process look like — a skill is opinionated middleware, not a neutral capability. He proposes a “one-job test”: run a real task through the skill across seven steps and decide to keep, rebuild, or delete it, treating skills as testable tools that require validation rather than passive installs.

Implications

Feeds the agentic-engineering radar thread on context economics and the “more isn’t better” pattern. Pairs with the companion signal that 25 skills produce duller work than 5 — both frame skill proliferation as a cost, not a free accumulation. Rhymes with the harness-side moves already tracked (CC’s “medium” <15-agent workflow-size guideline; the ARC-AGI harness-config-as-multiplier finding): the field is learning that agent quality is bounded by what you don’t load as much as what you do. A skill is provenance-of-taste — whose definition of “good” is executing when you didn’t write it.