DEV Community

Thomas Lau
Thomas Lau

Posted on

Cut Claude Code Token Usage by Delegating to Cheaper Models with Boss Mode

#claude #llm #productivity #devtools

boss-mode

boss-mode banner

A Claude Code skill that stops Claude from spending tokens on work a cheaper model can handle.

Claude acts as the orchestrator (boss): it plans, delegates, reviews, and verifies.

Cheaper models act as workers: they write the code, drafts, and analysis.

The Three-Tier Model

Tier Who When
Boss Claude Planning, architecture, file ops, final decisions, talking to you
Senior worker DeepSeek (or similar) Code drafts, reasoning, analysis, test generation
Bulk worker Groq / Ollama Translation, formatting, high-volume repetitive tasks

The idea: Claude's tokens are expensive. Only spend them where Claude's judgment is actually needed.

The Boss Loop

Every task follows this cycle:

1. PLAN    — Claude breaks the request into sub-tasks, picks the right worker for each
2. DELEGATE — Claude sends a structured prompt to the worker
3. REVIEW  — Claude reads the output, accepts or requests a targeted fix
4. VERIFY  — Claude runs tests/checks; if they fail, loop back to DELEGATE
Enter fullscreen mode Exit fullscreen mode

Delegation Protocol

Every delegation prompt uses this template — workers have no file access, so include all context inline:

CONTEXT: [project background, relevant code snippets, existing patterns]
TASK:    [exactly one deliverable — be specific]
CONSTRAINTS: [style rules, what NOT to change, existing patterns to follow]
SUCCESS: [a test command or checklist that proves the output is correct]
Enter fullscreen mode Exit fullscreen mode

Example — delegating a Python module to DeepSeek:

CONTEXT: Python project using SQLite + akshare. Existing income.py fetches income
         statements and saves rows to a 'income_statement' table. Pattern attached below.
         [paste relevant income.py snippet]

TASK: Write balance.py that fetches balance sheet data via
      akshare.stock_balance_sheet_by_report_em(symbol), saves rows to SQLite
      table 'balance_sheet' with the same column naming convention as income.py.

CONSTRAINTS: Follow income.py exactly — same error handling, same save() pattern,
             same logging style. Do not add new dependencies.

SUCCESS: pytest tests/test_balance.py passes. Table 'balance_sheet' created with
         correct schema (verify with: sqlite3 data.db ".schema balance_sheet").
Enter fullscreen mode Exit fullscreen mode

Claude sends this to the worker, receives the draft, writes it to disk, and runs the tests.

Installation 

# 1. Clone
git clone https://github.com/thomaslau0229/boss-mode.git
cd boss-mode

# 2. Run the setup script (installs + configures worker paths)
bash setup.sh

# 3. Restart Claude Code
Enter fullscreen mode Exit fullscreen mode

The setup script will ask for your worker command paths and patch them into the skill automatically. You can also install manually by copying SKILL.md to ~/.claude/skills/boss-mode/SKILL.md.

Claude Code scans skill descriptions at startup. It will automatically load boss-mode when it detects a task that fits the trigger conditions.

Configuration — Wire Up Your Workers

boss-mode expects you to have worker scripts or MCP tools Claude can call via Bash. Set these up once:

Option A: Script-based workers

Create a simple wrapper for each model tier. Example for DeepSeek:

# ~/.claude/workers/deepseek_worker.py
import sys, os, requests

prompt = sys.argv[1]
resp = requests.post(
    "https://api.deepseek.com/chat/completions",
    headers={"Authorization": f"Bearer {os.environ['DEEPSEEK_API_KEY']}"},
    json={"model": "deepseek-chat", "messages": [{"role": "user", "content": prompt}]}
)
print(resp.json()["choices"][0]["message"]["content"])
Enter fullscreen mode Exit fullscreen mode

Then update the "Calling Workers" section in SKILL.md to point to your actual script paths.

Option B: MCP tools

If you have DeepSeek or Groq set up as MCP tools in Claude Code, reference the tool name directly in SKILL.md instead of a Bash command. Claude will call it natively.

Environment variables

Store your API keys in your shell profile, never in the skill file:

export DEEPSEEK_API_KEY="your-key-here"
export GROQ_API_KEY="your-key-here"
Enter fullscreen mode Exit fullscreen mode

Auto-Trigger Scenarios

Once installed, Claude applies boss-mode automatically when it sees:

  • About to write >30 lines of new code → delegates draft to DeepSeek
  • Executing a superpowers implementation plan → delegates each coding sub-task
  • Writing test suites or test cases → delegates to DeepSeek
  • Writing documentation, PR descriptions, commit messages (>200 words) → delegates draft
  • Debug reasoning: analysing error logs, tracing root cause → delegates reasoning to DeepSeek
  • Batch repetitive tasks (translation, formatting) → delegates to Groq/Ollama
  • Pros/cons analysis or option comparison → delegates to DeepSeek, Claude makes final call

You can reinforce this by adding boss-mode to the skill table in your CLAUDE.md:

| Writing code / fixing bugs  | **boss-mode** → superpowers:systematic-debugging → superpowers:verification-before-completion |
| Multi-step implementation   | **boss-mode** → superpowers:writing-plans → superpowers:executing-plans |
Enter fullscreen mode Exit fullscreen mode

Integration with Superpowers

boss-mode is designed to sit in front of superpowers workflows, not replace them:

Superpowers skill boss-mode role
writing-plans Claude writes the plan (no delegation — this needs Claude's judgment)
executing-plans Each coding task in the plan → delegate to worker, Claude reviews
brainstorming Claude brainstorms; deep analysis sub-tasks → delegate to DeepSeek
verification-before-completion Always Claude — never delegated
systematic-debugging Claude diagnoses root cause, delegates fix writing

Fix Loop

When a worker's output fails tests:

1. Read the exact error — don't over-diagnose
2. Write ONE targeted fix instruction
3. Delegate to the same worker
4. Re-run the test
5. Repeat up to 3 rounds
6. After 3 rounds: Claude writes the fix directly and notes why the worker struggled
Enter fullscreen mode Exit fullscreen mode

What Claude Never Delegates

  • Reading or writing files
  • Running shell commands or tests
  • Final architecture decisions
  • Responses directly to the user
  • Security-sensitive logic

Acknowledgements

boss-mode is built on top of patterns and frameworks from the following open-source projects:

  • superpowers — The skill framework and workflow patterns (writing-plans, executing-plans, verification-before-completion, etc.) that boss-mode integrates with and extends.
  • Andrej Karpathy's LLM coding guidelines — The four principles (Think Before Coding, Simplicity First, Surgical Changes, Goal-Driven Execution) that inform how Claude delegates and verifies work. Formalised as a Claude Code skill at andrej-karpathy-skills.
  • agentskills.io — The skill specification format used by SKILL.md.

Contributing

Issues and PRs welcome — especially new worker templates, delegation prompt examples, or integration patterns with other Claude Code skills.

License

MIT

中文说明

一个 Claude Code 技能，让 Claude 不再把昂贵的 token 浪费在便宜模型也能干的活儿上。

Claude 扮演决策者的角色：负责规划、分工、审查和验证。

更便宜的模型扮演员工：负责写代码、起草文档、做分析。

三层模型

层级 谁 什么时候用
决策者 Claude 规划、架构设计、文件操作、最终决策、跟你沟通
高级员工 DeepSeek（或类似模型） 代码起草、推理分析、测试用例生成
执行员工 Groq / Ollama 翻译、格式化、大批量重复性任务

核心理念：Claude 的 token 很贵。只在真正需要 Claude 判断力的时候才用它。

决策者工作流

每个任务都走这个循环：

  1. 规划 — Claude 把需求拆成子任务，给每个子任务分配合适的员工
  2. 委派 — Claude 给员工发送结构化的 prompt
  3. 审查 — Claude 读输出，通过就接受，不通过就要求定向修改
  4. 验证 — Claude 跑测试/检查；失败了就回到「委派」那步重来

委派协议模板 

CONTEXT: [项目背景、相关代码片段、现有模式]
TASK: [明确的一件事 —— 要具体]
CONSTRAINTS: [风格规则、哪些不能改、要遵循的现有模式]
SUCCESS: [能证明输出正确的测试命令或检查清单]
Enter fullscreen mode Exit fullscreen mode

安装 

git clone https://github.com/thomaslau0229/boss-mode.git
cd boss-mode
bash setup.sh   # 自动安装并配置 worker 路径
Enter fullscreen mode Exit fullscreen mode

自动触发场景

安装后，Claude 会在以下情况自动启用 boss-mode：

  • 即将写超过 30 行新代码 → 把草稿委派给 DeepSeek
  • 写测试套件或用例 → 委派给 DeepSeek
  • 写文档、PR 描述、commit 信息（超过 200 字）→ 委派草稿
  • 调试推理：分析错误日志、追溯根因 → 委派给 DeepSeek
  • 批量重复任务（翻译、格式化）→ 委派给 Groq/Ollama

致谢

Top comments (0)