Agentfile Hygiene Rubric

REFERENCE MODE: This skill provides reference material. Load the rubric on-demand when auditing, trimming, or gating an agentfile.

An agentfile is an always-on instruction file an AI coding assistant loads into context every session — CLAUDE.md (Claude Code) and AGENTS.md (OpenAI / OpenCode spec). Because the file is always-on, every line carries a fixed per-session token cost and competes for the context window; agentfiles drift upward over time as rules accrue. This skill carries the objective, domain-invariant rubric for deciding which sections earn that always-on cost.

The rubric is the single normative source of truth consumed by BOTH halves of the agentfile-hygiene capability:

• the cognitive sweep plan-marshall:recipe-agentfile-hygiene, which classifies each section and emits one remediation deliverable per offending section; and
• the deterministic pm-plugin-development:plugin-doctor backstop rules agentfile-line-count-over-budget and agentfile-directory-tree-present.

Consumers reference the rubric — they MUST NOT restate its content — so the LLM recipe and the deterministic rules stay in sync by construction.

Enforcement

Execution mode: Reference library; load the rubric standard on-demand.