Your agents are dirty, inefficient, and wasting resources.
Clean your agents -> Improve your habits -> Save the world 🌍
LeanMaxxing Leaderboard soon?
npx agent-hygieneLet agents help you clean w/ exact prompts.
22 checks across four categories. Auto checks read your config files. Session checks analyze real usage via AgentsView. Advisory checks are habit reminders.
| Check | Tier | What it catches |
|---|---|---|
.claudeignore coverage |
Auto | Build artifacts flooding context |
CLAUDE.md size |
Auto | System prompt over 80 lines |
AGENTS.md size |
Auto | Codex instructions over 80 lines |
| Rules verbosity | Auto | CLAUDE.md/AGENTS.md written with bulleted bloat and filler |
| Skills vs fat CLAUDE.md | Auto | Domain knowledge crammed into CLAUDE.md instead of skills |
| MCP tool search deferral | Auto | All MCP tools loaded upfront instead of deferred |
| Context window pressure | Session | Peak session exceeding 100K tokens |
| Check | Tier | What it catches |
|---|---|---|
| Autocompact threshold | Auto | No compaction until context is already huge |
| Subagent model override | Auto | Subagents running on Opus instead of Haiku |
| Default model selection | Auto | Opus set as default model |
| Opus model overuse | Session | >40% of tokens going to the priciest model |
| Prompt cache efficiency | Session | <30% cache hit rate |
| Subagent cost efficiency | Session | Subagents burning Opus tokens |
| Check | Tier | What it catches |
|---|---|---|
| Path-scoped rules | Auto | No .claude/rules/ directory |
| Settings JSON schema | Auto | Missing $schema in settings.json |
| Merge tiny rule files | Auto | Too many small rule files that should be combined |
| Codex configuration | Auto | Missing or misconfigured Codex setup |
| Check | Tier | What it catches |
|---|---|---|
/clear between tasks |
Advisory | Context bleeding between unrelated tasks |
/btw for side questions |
Advisory | Breaking flow for quick lookups |
| Default effort level | Advisory | Running at max effort when low/medium works |
| Delegate to subagents | Advisory | Doing research in the main thread |
| Plan mode for complex work | Advisory | Jumping into multi-file changes without /plan |
All 18 agent-CLI techniques documented here: docs/techniques/. Techniques for direct API/SDK use live under docs/api-direct-usage-techniques/.
scan- letter grade (A+ to F) with category breakdown and per-check detailsscan --fix- auto-fix failing checks where possiblesnapshot save/compare- before/after baselines with cost deltastrack- token usage trends, model cost breakdown, cache hit rates (via AgentsView)profile- all-time stats and blind spots (checks that have never passed)history- rolling score history with sparklineexport- anonymized JSON for leaderboard sharing (shipping soon)badge- shields.io-style SVG for your repo
Want to use Agent Hygiene in a GitHub Action? See an untested example in action.yml.
4 of the 22 checks (the Session tier) rely on real token usage data from your agent sessions. AgentsView provides that data- it reads your local agent logs and calculates per-session costs, model breakdowns, and cache hit rates without sending anything off your machine.
Without AgentsView installed, those 4 checks get skipped and you'll see a notice in the scan output.
To install:
curl -fsSL https://agentsview.io/install.sh | bashOn Windows:
powershell -ExecutionPolicy ByPass -c "irm https://agentsview.io/install.ps1 | iex"Once installed, run agent-hygiene scan again and the Session checks (Opus overuse, context bloat, cache miss rate, session length, subagent cost) will light up.
TokenMaxxers can't have all the fun.
Claude Code, Codex
Want to add checks for Cursor, Windsurf, Copilot, or another agent? There's a step-by-step prompt that walks your AI through the entire process- discovery, checks, tests, and docs. See contrib/add-agent-checks/PROMPT.md.
- Thanks to @atre for the inspiration and the original techniques behind this project. Read his article: Claude Code in Production
- Thanks to @tkmx for the inspiration for the leaderboard.
MIT - Copyright (c) 2026 Patrick Gallagher



