DEV Community: Thomas Lau

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

Thomas Lau — Wed, 29 Apr 2026 15:42:27 +0000

boss-mode

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

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]

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").

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

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"])

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"

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 |

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

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：

规划 — Claude 把需求拆成子任务，给每个子任务分配合适的员工
委派 — Claude 给员工发送结构化的 prompt
审查 — Claude 读输出，通过就接受，不通过就要求定向修改
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]

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

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

superpowers — 技能框架与工作流模式
Andrej Karpathy 的 LLM 编码指南
agentskills.io — 技能规范格式

I Had Zero Coding Knowledge. Claude Code Helped Me Build the Tech Behind My Drone E-Commerce Store.

Thomas Lau — Mon, 27 Apr 2026 10:14:37 +0000

I'm Thomas Lau. A year ago, I couldn't read a line of code. Today I run Reboot Hub — a Hong Kong-based e-commerce store specialising in Grade A+ Pre-Owned DJI drones and chip-level drone repair — and I built most of the technical infrastructure behind it with Claude Code.

This is the story of how that happened.

The Business Came First

I started with the drone expertise, not the tech. I'd been working in the drone industry in Hong Kong for years, and I saw a gap: DJI releases new flagship models every 12-18 months, flooding the secondhand market with perfectly good drones that nobody trusted. And when drones broke, official repair was expensive and slow.

So I got our team MOHRSS Level 3 certified — China's highest national standard for drone technicians — and started Reboot Hub to do two things: sell grade-verified pre-owned DJI drones with real warranties, and do proper chip-level board repairs.

The business made sense. The website was the problem.

"Just Use Shopify" Only Gets You So Far

I set up Shopify. I listed products. I wrote descriptions by hand.

Then I looked at our SEO situation: 285 products with inconsistent titles, no schema markup, no internal linking between articles and products, missing meta descriptions across the board. I had blog content but it wasn't connected to anything.

A friend told me to hire an SEO agency. The quotes were HKD 15,000-30,000 per month. For a bootstrapped startup, that wasn't an option.

Finding Claude Code

I'd been using Claude for writing. Then I found out about Claude Code — the CLI that can actually read your codebase, write scripts, and execute them.

I didn't know Python. I didn't know what the Shopify Admin API was. I opened Claude Code, described my problem, and asked: "Can we fix this without me understanding how to code?"

The answer was yes. But the more accurate answer was: by the end of this, you'll understand more than you think.

What We Built Together

Over the following weeks, Claude Code and I built a set of Python scripts that automated the SEO work I couldn't do manually:

Batch product updates. We wrote a script that fetched all 285 products via the Shopify API, standardised the title format to our brand spec ("Grade A+ Pristine Pre-Owned DJI [Model]"), and updated meta descriptions at scale. What would have taken weeks of clicking took one script run.

Schema injection. Google and AI crawlers need structured data to understand what a site is. We wrote a script that injected Organization JSON-LD schema — including our MOHRSS certification and social profiles — directly into the Shopify theme file via the Assets API. No app required.

Internal linking. Every time we published a blog article about, say, DJI Air 3S repair, relevant keywords in that article should link to our pre-owned Air 3S collection. We built a script that did this automatically across the entire blog.

Multilingual content. We serve customers in Russia and Portugal. Claude Code helped us build a translation pipeline using local LLMs — no paid API required — to translate product titles, meta descriptions, and full article bodies into Russian and Portuguese.

AI crawler discovery. We created a llms.txt file at reboot-hub.com/llms.txt — the emerging standard for helping AI systems like ChatGPT and Perplexity understand your site structure. On Shopify this required a creative workaround (custom page template + redirect), which Claude Code figured out.

What I Actually Learned

I still can't write Python from scratch. But I can now:

Read a script and understand what it does
Spot when something looks wrong
Ask Claude Code the right questions to debug
Understand why a particular approach matters (rate limits, idempotency, cursor pagination)

The gap between "I can't code" and "I can build things with AI" turned out to be much smaller than I thought. What Claude Code provided wasn't just code — it was the ability to have a technical conversation with someone who understood both what I was trying to build and how to build it.

The Honest Part

It's not magic. There were scripts that ran wrong and needed fixing. There were times I didn't understand why something wasn't working and had to describe the problem three different ways before Claude Code understood what I meant.

But for a non-technical founder running a bootstrapped business, the alternative was either paying agency rates I couldn't afford, or leaving the SEO work undone. Claude Code made a third option real.

Where Reboot Hub Is Now

The store is live at reboot-hub.com. We sell Grade A+ Pre-Owned DJI drones with 180-day warranties, and we run a global mail-in repair service from Hong Kong — customers worldwide ship us their broken DJI drones, we fix them at the board level, and ship them back.

The SEO pipeline runs daily. The schema is in place. The internal linking is done.

And I didn't have to learn to code to build any of it.

If you're a non-technical founder who's been told you need to hire someone to do this — I'd push back on that. Happy to talk through what worked and what didn't in the comments.