Guidelines and Documentation Standards for OpenClaw Agent Development

Sourcing Inspiration for OpenClaw Agents

The primary requirement for agent creation is a verified source of inspiration from public platforms such as X, Twitter, Blue Sky, Reddit, or professional blogs and magazines like Medium or Outlet. Valid sources must show real-world discussions or implementations of OpenClaw, OpenClot, or OpenCLOA by users other than the contributor. Contributors are prohibited from using generic links, internal ChatGPT conversation ideas, or general healthcare risk stratification descriptions that do not reference a specific online post. If an idea is original, the contributor may post it themselves publicly first to provide a valid URL and screenshot for the record.

Documentation and Retrieval Requirements

Every task must provide the source name, the full URL link, a screenshot of the post to prevent loss due to potential deletion, and the exact retrieval date. When a link contains multiple ideas, such as a Reddit thread or Medium article listing 3030 use cases, the source details section must specify the exact idea being used. This level of detail ensures that customers and reviewers can pinpoint the specific inspiration point during quality control and verify that the agent is grounded in actual human use cases for OpenClaw.

Defining the Agent Objective

The agent objective must provide comprehensive context regarding the background, goals, and desired outcome of the task. It should summarize the involved files—such as emails, CSVs, or Google Drive documents—and highlight environmental challenges like duplicate records, irrelevant information, or incomplete data. Descriptions must be sufficiently detailed so that an unfamiliar reviewer can understand the workflow, such as managing a wine cellar or a news blog, and evaluate whether the model successfully navigates the provided datasets to produce a high-quality result.

Case Study: Wine Cellar Management

In the wine cellar manager example, the agent handles a collection of 1010 wines and 1414 bottles across multiple storage locations. The task involves reading files like wine cellar CSV, Vivino export CSV, and reference notes from Google Drive. The agent must reconcile inconsistent data, identify drinking windows, and produce an intelligence report. Specific success metrics include identifying 77 bottles to drink now, 22 to hold, and 11 bottle past its peak. It may also involve complex tool interactions such as lookup tasks from Wikipedia, saving draft emails in Gmail, and setting reminders in Google Calendar.

Case Study: News Content Pipeline

For single-turn tasks, the news blogger agent example demonstrates how a model manages a content pipeline within a Google Drive document. The agent must ingest headlines and prioritize a queue based on an editorial capacity limit of 55 published stories per day. The workflow requires the model to normalize entries, merge duplicate stories, and compute priority scores while providing justification for excluded content. The final output must be a structured JSON publication queue, emphasizing the importance of defining precise output formats to facilitate auditability and rubric-based evaluation.