Doubao Work Zero-to-One: A Step-by-Step Agent Practical Guide for Beginners
ByteDance has officially integrated its AI office ecosystem: combining TRAE, Coze, Doubao LLMs, and native Feishu / Lark integration into a unified desktop entry — Doubao Work (豆包工作).
Unlike standard chatbots that simply converse in text, Doubao Work breaks down objectives into multi-step execution plans, manipulates local files, connects to external tools, and outputs tangible, editable deliverables (Word, Excel, PPTX, Lark Base, images).
9 Core Navigation Portals
Single-shot execution portal for ad-hoc document and research tasks
Cron-like background recurring runs for weekly reports and monitoring
Auto-aggregates local .codex/skills (200+ pre-existing tools)
External API integrations: Lark, DingTalk, WeChat Work, GitHub, Notion
Persistent specialized personas (Data Analyst, PPT Expert, UI Designer)
Multi-agent pipeline coordination across research, writing, and design
Direct read/write access to Lark Cloud Docs and Base tables
Monitor and instruct desktop agent tasks on the go via smartphone
Long-term context repositories with directory sandboxing
Key Practical Prompts
读取【材料文件】。 先给我一份 8 页 PPT 页纲，确认后再生成可编辑演示稿。 内容包含：背景、三个关键发现、问题、建议和下一步。 不要编造材料里没有的数据；引用数字时标明来源页或表格位置。 保存为【输出文件夹】/项目汇报_v1.pptx，不覆盖原文件。 验收：共 8 页，每页只有一个中心结论，所有数字能回到原材料。
The Fundamental Golden Rule:
"Let the Agent execute first, then human verifies the deliverables."