Building a Google Prompt‑Engineering Assistant with Coze

Recognize that writing prompts is inherently counter‑intuitive because prompts must be phrased for machines rather than humans, similar to describing a scene to a drawing AI with detailed, technical terms.

Google’s Prompt‑Engineering Whitepaper provides systematic, in‑depth guidance on effective prompt creation, covering LLM output parameters (Top‑K, Top‑P, Temperature), prompt techniques (zero‑shot, few‑shot), and best practices (conciseness, clarity).

To make this knowledge easily accessible, a Coze assistant called Google Prompt Assistant was built, which stores the whitepaper in a knowledge base and uses a large model to answer queries and guide users in writing high‑quality prompts.

The assistant works in two steps:

Upload the entire whitepaper into a Coze knowledge base.

When a user inputs a query, the large model retrieves relevant content from the knowledge base and generates a response.

Creating the knowledge base involves navigating Coze’s workspace, selecting the resource library, adding a new resource, uploading the PDF, and completing the creation wizard.

After the knowledge base is ready, a workflow is built to classify user intents into three categories: (1) document‑related queries about prompt engineering, (2) requests for prompt generation guidance, and (3) unrelated queries (e.g., weather). Each branch uses the DeepSeek V3 model and references the knowledge base.

The workflow is linked to a Coze agent, enabling two main use cases: querying prompt‑engineering knowledge and receiving step‑by‑step prompt‑writing assistance.

Example prompts and their optimized versions from the whitepaper are shown, illustrating why certain phrasing yields better results.

Additional resources include download links for the bilingual whitepaper, example usage scenarios, and a link to the live Coze assistant.

Overall, this tool aims to teach users the techniques behind perfect prompts rather than simply providing static prompt examples.