agent-harness-engineer

Start with a read-only mapping pass that inventories every instruction file, skill, memory store, and permission, then evidence-label each rule and apply the six-outcome triage: keep, one-home, load-later, turn-into-check, probation, or retire. Changes ship as numbered, reversible proposals.

Common causes are duplicated rules drifting across files, binary requirements buried in prose, and too much material loaded at startup. The fix is one canonical home per rule, moving yes/no guarantees into validators or hooks, and loading specialist material only at the phase that needs it.

Design a knowledge layer: define the work object, retrieval unit, authoritative source, permissions, provenance, and precompiled context in a Retrieval Contract Spec before choosing technology. Label all write-back so model guesses never silently become instruction.

Test both with the 3x3 method: run the same job three times with a compact brief and three times with the full method, score content blind, and check the binary delivery contract separately. Practitioner reports show long context can improve analysis while degrading delivery when enforcement stays in prose.

Copying a SKILL.md between tools fails because loading behavior, triggers, hooks, and permissions differ. Build a portable work package covering install location, trigger behavior, dependencies, memory scope, and tested harnesses, generated from one markdown source.

Low-risk assistants do not need the full knowledge layer, and most sessions do not deserve a new skill. Default to the smaller scope: one canonical rule home and a few prose-to-check conversions beat an unfinished grand redesign.