Build an AI role someone else can use through chat

OfficeCubs offers Crew for your own desktop tasks and Familiar Chat for helping someone else through their messaging app. This guide covers Familiar Chat.

Familiar Chat connects two kinds of people: the developer who knows how to configure AI tools, and the person who wants to ask a question in a familiar chat app. The developer prepares a role with Codex, Claude Code or DeepSeek Harness. The other person talks to that role through a supported private conversation in WeChat, Feishu / Lark or Telegram.

Define a role around the user’s question

Start with the person you are helping. A colleague may need answers from a product guide. A family member may want help understanding a supplied document. A client may need drafts in an approved voice. Give the role a narrow purpose, suitable source material and clear instructions for missing information.

The role and the work engine are separate choices. A product knowledge role can be configured to use Codex or Claude Code; the developer chooses the tool behind that teammate. This is not a promise that different engines behave identically or automatically run together.

Configure and test on the host computer

1. Install and authorize a supported AI tool on the computer running OfficeCubs, then check it in Office settings.
2. Create or recruit the role. Write its working guide, bind relevant skills and select the workspace or Docker sandbox.
3. Choose the work engine for that teammate. Keep provider credentials in the supported authentication setup, outside the role’s instructions.
4. Ask one question with a known answer and one whose answer is absent from the supplied sources. Review both before connecting another person.

A working guide you can adapt

You help a nontechnical colleague understand our approved product notes. Answer in plain language, name the source document and preserve important limitations. If the notes do not answer the question, say what is missing and ask a focused follow-up. You may draft a message when asked, but do not claim it has been sent.

Replace the sample subject with the user’s actual need. Supply the reference files on the host; a prompt referring to missing documents cannot make them available.

Connect the intended person, not just the developer

For Telegram, prepare a dedicated bot and help the intended user pair their private conversation. For Feishu / Lark, finish the supported organizational authorization and allowed-sender setup. For WeChat, the intended user must scan and confirm the connection QR code with their own account. The account that scans is the account permitted to start work.

The user then asks a normal question in chat. They do not need to install a CLI, choose a model or edit the role’s system instructions. The connected teammate runs the request through its selected AI tool on the host and returns a reply through that messaging connection.

Review the whole handoff

Have the user ask the two test questions from their own chat. Verify that the intended teammate receives the task, that its answer reflects the supplied sources, and that the reply reaches the correct conversation. Ask the user whether the answer makes sense without knowledge of the setup screens.

Keep the operating limits clear

The host computer and OfficeCubs must stay awake and online. The operator handles any tool approval requests in the desktop app. Provider access and usage charges belong to the configured AI account; task content may reach the model provider.

These are currently private-chat integrations. Telegram, Discord and Slack each support one paired private chat per provider connection in an office; this is not an unrestricted multi-user bot service. WeChat availability depends on its ClawBot rollout, and Feishu requires the appropriate provider configuration. File delivery varies by provider. Start with a text question and review important answers before relying on them.