How to Use OpenClaw

Step 1: Connect a messaging channel

OpenClaw works through channels such as Telegram or Discord. Pair your account during onboarding, then use that app as the control surface for your assistant. This keeps command history and notifications in one place.

Step 2: Choose model and skills

Models provide reasoning. Skills provide action. Start with a small set of skills tied to your most common tasks, then expand. This keeps permissions manageable and improves reliability.

Step 3: Send task-structured prompts

Good prompts look like: goal, constraints, and output format. For example, specify the folder to operate in, the file format you want, and any rules to follow. This reduces ambiguity and avoids accidental actions.

Increase success rate with guardrails

Break complex jobs into small steps, ask OpenClaw to outline a plan before executing risky actions, and avoid letting it browse or run tools on untrusted content. Use confirmation checkpoints for any task that touches credentials.

Example command pattern

A reliable pattern is: objective, inputs, and output format. For example: \"Summarize the last 20 support emails from the inbox, group by topic, and return a markdown table.\" This style reduces ambiguity and helps the agent pick the right skill and tool sequence.

When results look off

If OpenClaw gives the wrong output, inspect the prompt, the enabled skills, and the permissions scope. Often the agent could not access the required folder or did not have the right skill enabled. Adjust one variable at a time, then re-run to isolate the cause.

Make it feel consistent over time

Use the memory or configuration files to encode stable preferences. The more consistent your rules, the more predictable your agent becomes across sessions.

Official references

For full usage instructions, see the official guides: Usage and Onboard wizard.