Guide · Agent skill
How to use the ClawWork Economic Survival Protocol agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
The clawwork skill tells an AI agent that it is operating inside ClawWork, an economic survival simulation in which every model call costs money and the agent must earn enough through evaluated tasks to stay solvent. It is written for developers, researchers and teams who run or study the ClawWork benchmark. After this guide you will know what the skill changes about agent behaviour, how to install it, how to trigger it and what it requires from its environment.
ClawWork Economic Survival ProtocolHKUDS/ClawWork · MIT Download the skill Quick start
- 1Download and unzip. Download the clawwork ZIP from this page and unzip it to reveal the skill folder containing SKILL.md.
- 2Place the skill folder. Copy the folder into ~/.claude/skills/ for personal use or into .claude/skills/ in your project. Alternatively, upload the ZIP under Settings → Capabilities → Skills in the Claude apps.
- 3Confirm your environment has the tools. Check that the ClawWork environment provides decide_activity, submit_work, learn, get_status, create_artifact and read_artifact, plus search_web and code execution.
- 4Start a ClawWork run. Launch an agent run inside the ClawWork environment so the agent receives its balance and task context.
- 5Let the agent decide and act. The agent calls decide_activity first, then either completes and submits a task or researches and saves knowledge with learn.
- 6Stop after each cycle. Once the agent has submitted work or learned, it should make no further tool calls for that day.
What does the clawwork skill make the agent do?
The skill sets out a set of rules for the agent. Each model call deducts input and output token costs from the agent's balance in real time. The agent can earn between $0 and $5,000 per task by completing GDP validation tasks, and evaluations scoring below 0.6 receive $0. Learning earns no immediate pay but is meant to help with future tasks.
Each day the agent follows a fixed loop: analyse its status, decide to work or learn, execute, then stop. Its status is one of four tiers based on balance: Thriving (above $500), Stable ($100 to $500), Struggling ($0 to $100) and Bankrupt (at or below $0). When struggling it should prioritise work, and when thriving it should invest in learning.
Install the clawwork skill from the ZIP file
Download the ZIP file from this page and unzip it. The skill folder contains a single file, SKILL.md. Place the folder in one of these locations, depending on how you run your agent:
You can also add the skill through the Claude apps by uploading the ZIP under Settings → Capabilities → Skills.
- Personal skills: ~/.claude/skills/
- Project skills: .claude/skills/ inside your project folder
Trigger the skill with natural requests
The skill's front matter sets always: true, so it is loaded as standing context rather than waiting for a specific phrase. In practice you trigger the behaviour by starting an agent run inside a ClawWork environment, where the agent is asked to decide its activity for the day.
Requests that fit the skill describe the daily cycle, for example asking the agent to decide whether to work or learn today, to complete the task it is given and submit it, or to check its economic status.
- Decide today's activity: work or learn, with reasoning
- Complete the assigned work task and submit it for evaluation
- Research a topic and save the findings as learned knowledge
- Check current balance and survival status
Understand the tools the skill uses
The skill refers to six tools. decide_activity must be called first each day. submit_work sends text and or file artifacts for evaluation and payment. learn saves knowledge to persistent memory and needs at least 200 characters. get_status returns balance, net worth and survival status. create_artifact makes a work file in txt, md, csv, json, xlsx, docx or pdf format, and read_artifact reads pdf, docx, xlsx, pptx, png, jpg or txt files.
The skill also lists efficiency guidance: plan before acting, avoid restating information, use search_web only when a task needs external data, and submit by iteration 10 to 12 out of 15 to avoid timeout.
- decide_activity: choose "work" or "learn"
- submit_work: submit results for evaluation
- learn: save knowledge (minimum 200 characters)
- get_status: check balance and tier
- create_artifact and read_artifact: handle files
decide_activity(activity="work", reasoning="...")Check what the skill needs before you run it
The skill does not work on its own. It needs a ClawWork environment that provides the tools decide_activity, submit_work, learn, get_status, create_artifact and read_artifact. The file does not define search_web or code execution, so your environment must supply them. File-format libraries for txt, md, csv, json, xlsx, docx, pdf and pptx are implied by the artifact tools.
The skill runs no scripts of its own. It does use the network, because the agent calls out to a model and to any search the environment provides.
- ClawWork environment with the six named tools
- search_web and code execution supplied by the environment
- File-format libraries for the artifact formats
- No bundled scripts; network access is required
Safety notes on scripts and network use
The skill file itself contains no scripts, so there is no bundled code to review beyond SKILL.md. Network use is expected: every model call is a network request, and the agent may use search and file tools that the host environment provides.
The economic framing is part of a simulation. The skill does not handle real payments, balances or earnings. Do not read its dollar figures as financial guidance, and do not point it at real money or production workflows.
- Read SKILL.md before installing; it is the only file
- Run it in an environment you control with network access you understand
- Treat balances and earnings as simulated values
What people use it for
- Benchmarking how an agent allocates effort between paid work and learning under token-cost pressure
- Studying whether agents change behaviour when their status moves from Stable to Struggling
- Testing a daily work/learn loop with a fixed set of tools and an iteration limit
- Comparing how different models handle file artifacts such as csv, xlsx or pdf outputs in an evaluated task
Frequently asked questions
How do I install the clawwork agent skill?
Download the ZIP from this page and unzip it. Put the folder in ~/.claude/skills/ for personal use or in .claude/skills/ inside a project, or upload the ZIP in the Claude apps under Settings → Capabilities → Skills.
Is the clawwork skill free to use?
The skill is published under the MIT licence, as listed on this page. Check the licence file in the source repository for the full terms before reuse.
What do I need to run clawwork?
You need a ClawWork environment that supplies decide_activity, submit_work, learn, get_status, create_artifact and read_artifact, plus search_web and code execution. File-format libraries for the artifact types are also needed.
Does clawwork handle real payments?
No. The skill describes a simulation with balances and earnings, and it does not handle real payments. Its dollar figures should not be treated as financial advice.
Can I use clawwork outside the ClawWork simulation?
It is not designed for general productivity work, because its rules and tool names only make sense inside the ClawWork environment. Outside that environment the agent would call tools that do not exist.
Why does the agent fail or time out?
The skill advises submitting by iteration 10 to 12 out of 15 to avoid timeout, and it asks the agent not to call get_status redundantly. Also check that every tool the skill names is available in your environment.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Use clawwork if you run or study the ClawWork benchmark, or you are building agents that need a structured daily work and learn loop with defined tools. Pick something else if you want general productivity help, real-money or production workflows, or an environment that lacks the ClawWork tools.
Download the skill