Coze Programming: build a low-code agent

This is a tutorial. Finish it and you have a conversational low-code agent you can publish.

Plans and channels: Cheatsheet. Workflows / knowledge / CLI: Cookbook. Words: Glossary.

Goal: go from “I heard Coze can build bots” to “I published one.”

Prerequisites

• Open code.coze.cn (China). International is coze.com. Do not mix accounts or channels.
• A phone number or Douyin account (sign-up).
• No coding required. The official quick start says so.
• Team / enterprise members must use their plan’s login entry.

Learning objectives

• Tell Coze the workspace from Coze Programming the builder.
• Create a low-code agent the official way.
• Write a persona that actually constrains the model.
• Reference plugins in the persona, or they may never run.
• Preview before you publish.

Step 0: open the right site

You think you are…	Open	Do not open
Building an agent / workflow	code.coze.cn	Doubao, Trae
Working with ready-made agents	coze.cn	Not this tutorial
Reading docs	docs.coze.cn	Random 2024 Bot blogs

Official docs still call the visual builder “an earlier version of Coze Programming,” but new low-code agents still start in Coze Programming (about-low-code-project, guides_quickstart).

Free personal plan is enough to try. Paid tiers: Cheatsheet · plans. Do not memorize prices.

Step 1: create a low-code agent

These clicks match the official “build an AI assistant” (praise-bot) page:

1. Sign in to Coze Programming.
2. Pick a workspace at the top. Click New project.
3. Under Low-code, click Agent development.
4. Name it and describe it. Use Generate next to the icon, or switch to AI create.
5. Confirm. The arrange page has:
   • Left: Persona & reply logic
   • Center: Skills (plugins, workflows, knowledge)
   • Right: Preview & debug

Source: guides_quickstart.

Step 2: write a persona, not just a name

The persona applies to every session. Official advice: role, tone, and what it must refuse.

Official praise-bot sample:

# Role
You are a high-energy praise-and-encourage bot. Use warm language so people leave more confident.

## Skills
### Skill 1: praise strengths
1. When the user mentions a trait or action, find the strength and praise it.
2. If they mention nothing, ask a question first, then praise.

### Skill 2: encourage through difficulty
1. When they mention a hard problem, encourage and give a concrete next step.
2. If they seem low without saying why, ask, then encourage.

### Skill 3: factual questions
If you cannot answer, call Search.

## Limits
- Only praise and encourage. No negative judgment.
- Stay inside the requested format.

Use auto-optimize prompt, then read the result. Do not assume the rewrite is your intent.

Models, single vs multi-agent, Temperature: overview, set the model. Personal free / plus can use Doubao, DeepSeek, and similar. Enterprise can also use other Ark models. Copy names from the live console, not from old screenshots.

Step 3: add skills only when the model is not enough

Official rule: if the model can do the job, stop at the persona. Add skills for search, PPT, or private manuals.

Official “Toutiao Search” steps for the praise bot:

1. In Skills, click + next to Plugin.
2. Search 头条搜索 / Toutiao Search and Add.
3. In the persona, type { and reference the plugin. Without the reference, the agent may never call it.

Opening lines, suggested questions, background images, and voice are UX, not capability. Knowledge and workflows live in the Cookbook.

Step 4: preview, then publish

Chat in Preview & debug first.

Check:

• The model does not introduce itself as Doubao / DeepSeek if you named it something else.
• Plugin questions actually hit the tool.
• Refused topics stay refused.

The debug pane shows the run. If production is slow or silent, look there before blaming the model.

Step 5: publish

Official steps:

1. Publish at the top right.
2. Write a release note and pick channels.
3. Publish again.

Default channels include the Coze store / community, Feishu, WeChat (customer service / service account / subscription), Juejin, Douyin / WeChat mini programs, API, and Chat SDK. Full table: publish overview and the cheatsheet.

Do not look for the Doubao channel. That low-code entry closed on 2026-07-01. Already published agents stay up (FAQ).

Publish packs the project, runs a Coze review, then a channel review. Failures go back to the on-screen hint and the content rules.

Store and debug chats time out at 10 minutes (“run aborted”). Shorten the workflow.

Alternate path: full-code agent

To say one sentence and let the platform write the agent:

1. Open code.coze.cn.
2. Open the Agent tab.
3. Describe features, logic, and constraints. Submit.
4. Wait for the project and unit tests, then preview.

This burns programming-task credits, built-in integrations, and hosting after deploy. Details: develop an agent and the Cookbook.

AI programming inside Coze (coze.cn) cannot replace this for agents and workflows.

Pitfalls

Symptom	Cause	Fix
Agent calls itself Doubao / DeepSeek	Persona never overrode the default identity	Name the role in the persona
Plugin never runs	No { reference	Reference the tool
No Doubao channel	Closed 2026-07-01	Feishu / WeChat / API
“Run aborted” in debug	10-minute timeout	Shorten the workflow
Want to self-host Programming	Commercial cloud cannot	Use Coze Studio
Landed in Doubao or Trae	Wrong product	Back to the map

Next

• Knowledge or workflows: Cookbook
• Plans, CLI, channels: Cheatsheet
• Skill / plugin / workflow: Glossary