Your first OpenClaw assistant should complete one small task you can verify. Start with a short research brief from text you supply, then inspect its sources and output before adding tools or automation.

This walkthrough operates an existing OpenClaw Gateway. It does not require a YAML project scaffold or a custom Python SDK. If installation is unfinished, follow the local quickstart.

Independent guide. Setup and workspace references reviewed on October 2, 2026. The sample task is a learning exercise, not a claim of a deployed research automation.

1. Confirm the assistant is ready

openclaw --version
openclaw gateway status
openclaw config validate
openclaw dashboard

Complete openclaw onboard if model access has not been configured. In the Control UI, send an ordinary text message first. A working reply gives you a baseline before file access or web research enters the task.

See the official quickstart for installation and service behavior. The earlier openclaw new, dev, and build commands on this page were not the documented workflow for operating this assistant and have been removed.

2. Give it a bounded input

Choose two or three short, non-sensitive paragraphs you are allowed to use. Paste them into a new conversation with a clear task:

Use only the text I provide below.
Create a brief with:
- Three supported findings.
- The sentence or paragraph supporting each finding.
- Two unanswered questions.
- A short list of next actions.
Do not browse, infer missing facts, contact anyone, or publish anything.
Mark uncertainty explicitly.

This exercise does not need web or shell access. It lets you evaluate the assistant's reasoning and reporting before granting more capabilities.

3. Check the result yourself

Compare every finding with the supplied text. Confirm the assistant separated supported observations from unanswered questions. If it invents a source or makes an unsupported conclusion, refine the instruction and repeat with the same input.

A useful result is one you can verify, not merely a fluent response. Save the input and a short note about corrections so you can compare the next attempt.

4. Add one capability only when needed

If you need the brief as a file, use a disposable workspace folder. Ask the assistant to save the reviewed result there only after confirming the file tools and policy available to that agent. Then inspect the actual file.

For web research, use the tools documented in the official tools overview. Ask for source URLs and dates, and open the important sources yourself. A returned link is not proof that its page supports the claim.

The beginner tools guide explains how to choose capabilities. Avoid installing a broad skill collection just to complete this one task.

5. Keep reusable instructions in the workspace

Inspect the selected workspace:

openclaw config get agents.defaults.workspace

Follow the official workspace conventions for reusable instructions. Keep the process short: purpose, allowed inputs, output format, and what to do when information is missing.

Runtime configuration belongs in JSON5 openclaw.json. The writing instructions above are plain text; they are not a made-up agents/research-assistant.yaml runtime schema. See the configuration guide before changing model or tool policies.

6. Test a missing-input case

Remove one paragraph or provide conflicting information. Check that the assistant marks the gap instead of inventing a resolution. If a required tool is unavailable, it should explain that limitation rather than claiming the task was completed.

Before adding a schedule or an external action, decide which outcomes need human review. Keep publishing, messaging, purchases, and destructive file changes outside this initial exercise.

7. Turn one successful task into a workflow

Once the brief works reliably with several sample inputs, document the steps in the workflow guide: input, decisions, tools, output, and review. Automate only the steps whose results you can check and whose permissions you understand.

For unattended availability, use the server guide. A personal Gateway is a service with state; it is not a static website created by openclaw build.

FAQ

Do I need to write code for this first assistant?

No. This exercise uses the Control UI and supported workspace conventions. Developing OpenClaw plugins is a separate task with its own official documentation.

Supplied text creates a stable, checkable input. Once that task works, adding web access introduces a specific new dependency you can evaluate separately.

How do I know a file was saved?

Inspect the requested path and contents. The assistant's statement that it wrote a file is not evidence of the filesystem outcome.