Stop wasting tokens. Start shipping. 6 battle-tested rules that cut Claude Code token usage by 50%+.
You use Claude Code daily. But have you noticed:
Round 1: Read file_a.md ← OK
Round 2: Read file_b.md ← should've been parallel
Round 3: Write file_a.html ← should've been parallel
Round 4: Write file_b.html ← 1 round wasted
Round 5: Re-read file_a.md ← was in system-reminder cache!
Round 6: Fix bug #1, restart training
Round 7: Fix bug #2, restart training ← should've collected all bugs first
Round 8: Fix bug #3, restart training
...
Round 15: Finally done. $12 in tokens. 40 minutes gone.
This isn't the model's fault. It's an execution problem. Claude doesn't know your efficiency preferences unless you tell it.
A drop-in CLAUDE.md ruleset that tells your agent how to execute, not what to do.
# 30 seconds. One command. Then forget about it.
curl -fsSL https://raw.githubusercontent.com/qingfengyu153781-star/claude-code-toolkit/main/scripts/install.sh | bashNext session, your agent automatically:
- Reads files in parallel
- Writes files in parallel
- Checks cache before re-reading
- Collects ALL errors before restarting
- Includes all 6 required elements in prompts
- Validates disk state before reporting
| Scenario | Before Toolkit | After Toolkit | Savings |
|---|---|---|---|
| Edit 10 HTML templates | 12-15 rounds | 2-3 rounds | ~80% |
| Launch ML training | 8 restarts, 2 hours | 1 restart, 15 min | ~87% |
| Generate image prompts | 3-5 iterations each | 1-2 iterations each | ~50% |
| Resume session after break | Re-survey entire project | Read cache, continue | ~30% |
| Fix environment issues | 10+ blind attempts | 2 rounds max | ~70% |
6 modular rules. Use all of them, or pick what you need.
| # | Rule | One-liner | Impact |
|---|---|---|---|
| 01 | Token Efficiency | Never re-read cached files; batch by default; diagnose before 3rd attempt | -40% |
| 02 | Batch Operations | N independent files = N parallel writes. Done in 2 rounds. | -60% |
| 03 | Session Management | Disk is truth. Check cache before reading. Validate before exit. | -30% |
| 04 | Prompt Checklist | 6-element self-check before ANY image/video prompt output | -50% |
| 05 | ML Training | API check first → venv first → all errors at once → one restart | -85% |
| 06 | Environment Setup | GPU compatibility pre-check → 2-round give-up → pip before compile | -70% |
User: "帮我批量改写10个HTML模板的配色方案"
│
▼
WITHOUT TOOLKIT WITH TOOLKIT
──────────────── ────────────
Read template_01 Read all 10 templates (parallel)
Write template_01 ┃
Read template_02 Establish 1 design pattern
Write template_02 ┃
... (×10) Write all 10 templates (parallel)
ls verify → Done
Total: 12-15 rounds Total: 2-3 rounds
Time: ~8 minutes Time: ~1 minute
- Heavy Claude Code users — if you run 50+ rounds/day, this saves $20-50/month
- Teams — consistent agent behavior across all team members
- Beginners — skip the "learn by wasting tokens" phase entirely
- Anyone who's ever thought "why did Claude just re-read that file?"
📊 "8 training restarts → 1" — LoRA fine-tuning session
8 sequential failures (HF symlink → transformers version → API change → dtype mismatch → ...), each requiring a restart that wasted 5-10 minutes of GPU idle time. With the toolkit: read full traceback, list all 8 errors, fix all at once, one restart. 2 hours → 15 minutes.
📊 "Prompt checklist eliminated our regeneration loops" — AI video production
Before: every image prompt was missing something — negative prompt, reference image path, model settings — requiring 3-5 iterations to get right. After: 6-element checklist enforced before output. 3-5 iterations → 1-2 iterations.
📊 "10 templates, 1 round" — Resume batch editing
Rewriting 10 HTML resume templates used to be 10 sequential operations. With batch rules: all 10 read in parallel, all 10 written in parallel, one ls to verify. 40 minutes → 3 minutes.
Your CLAUDE.md is the agent's "system prompt". When you add these rules:
Your CLAUDE.md
└── <!-- claude-code-toolkit START -->
├── Token Efficiency: "read cache before Read()"
├── Batch Operations: "N files → N parallel Write()"
├── Session Management: "ls before status report"
├── Prompt Checklist: "count 1→6 before output"
├── ML Training: "all errors → all fixes → one restart"
└── Environment: "2 rounds → ask user"
<!-- claude-code-toolkit END -->
The agent reads these at session start and follows them. No configuration, no plugins, no API keys. Just rules.
# Quick: one-command install
curl -fsSL https://raw.githubusercontent.com/qingfengyu153781-star/claude-code-toolkit/main/scripts/install.sh | bash
# Or: clone and pick what you want
git clone https://github.com/qingfengyu153781-star/claude-code-toolkit.git
cd claude-code-toolkit
python scripts/install.py --global # install to ~/.claude/CLAUDE.md
python scripts/install.py --project # install to ./CLAUDE.md
python scripts/install.py --all # both
# Or: zero-dependency manual install
# Copy claude.md → paste into your CLAUDE.md. Done.| Document | Language | Content |
|---|---|---|
| Quick Start | EN | 2-minute setup |
| Operation Manual | 中文 | Full guide with examples |
| Examples | EN | 5 real case studies |
| claude.md | — | The actual rules (what gets installed) |
Every rule in this repo exists because someone made a real mistake, then wrote a rule to prevent it.
Have a token-wasting pattern to report? Open an issue with:
- What happened (the mistake)
- How many rounds/tokens it wasted
- What rule would prevent it
Want to add a rule? Fork → add to rules/ → PR with before/after comparison.
- ⭐ Save money — every rule saves real API costs
- ⭐ Save time — fewer rounds = faster results
- ⭐ Help others — stars push this to more Claude Code users
- ⭐ It's free — MIT license, no strings attached
Built from 200+ hours of Claude Code.
Every rule has a scar. Every scar saved someone else.