ContextArchitect

对 Claude Code 和 OpenAI Codex 中的智能体上下文文件进行审计与优化。涵盖 CLAUDE.md、AGENTS.md、规则、技能及记忆组织。适用于“优化上下文”、“审计设置”、“上下文架构”、“该放哪里”、“减少重复”、“上下文预算”、“令牌过多”、“重新组织规则”、“技能层级”、“CLAUDE.md 太大”、“AGENTS.md 太大”或“优化技能”等场景。基于 Gloaguen 等人 2026 年的研究、Anthropic 最佳实践及 OpenAI Codex 文档。

name: ContextArchitect description: Audit and optimize agent context files across Claude Code and OpenAI Codex. Covers CLAUDE.md, AGENTS.md, rules, skills, and memory organization. USE WHEN "optimize context" OR "audit setup" OR "context architecture" OR "where should this go" OR "reduce duplication" OR "context budget" OR "too many tokens" OR "reorganize rules" OR "skill hierarchy" OR "CLAUDE.md too big" OR "AGENTS.md too big" OR "optimize skills". Based on Gloaguen et al. 2026 research, Anthropic best practices, and OpenAI Codex docs.

ContextArchitect

Analyzes your entire agent context setup — across both Claude Code (CLAUDE.md, .claude/) and OpenAI Codex (AGENTS.md, .codex/) — at user and project levels. Identifies duplication, bloat, misplaced content, and cross-agent inconsistencies.

Workflow Routing

Workflow	Trigger	File
Audit	"audit setup", "optimize skills", "check my organization"	`Workflows/Audit.md`
Placement	"where should this go", "user or project level"	`Workflows/Placement.md`

Context Files

•Principles.md — Research-backed principles (Gloaguen paper, Anthropic docs, OpenAI Codex docs, community patterns)

Examples

Example 1: Full audit

code

User: "audit my skill setup"
-> Scans ~/.claude/, ~/.codex/, and project-level files
-> Checks for duplication, bloat, misplaced content, cross-agent drift
-> Outputs prioritized recommendations with exact file moves

Example 2: Placement decision

code

User: "I have a code review workflow -- where should it live?"
-> Asks: project-specific rules? Used across all repos? Which agents?
-> Recommends user-level skill vs project-level rule
-> Shows exact directory path and template

Example 3: Cross-agent consistency check

code

User: "are my Claude and Codex configs in sync?"
-> Compares ~/.claude/CLAUDE.md with ~/.codex/AGENTS.md
-> Flags drift between the two
-> Recommends single source of truth pattern (stow-managed)