The Problem I Was Trying to Solve§

When asking a model to execute a complex task in a single prompt—such as researching a competitor, drafting a blog outline, writing SEO copy, and generating ad titles—the output quality drops. The model suffers from cognitive overload, forgetting specific constraints and introducing factual errors. I wanted to design a sequential automation pipeline that separates these concerns, passing variables from one prompt to the next to guarantee structured, professional outputs.

Tools and Setup§

For this pipeline, I configured a Node.js script calling the DeepSeek API. I utilized DeepSeek Chat for its low latency and precise JSON outputs.

The chain consisted of three components:

  1. Outline Creator: Converts seed concepts into an outline.
  2. Copywriter: Ingests the outline and writes the draft.
  3. Ad Planner: Takes the draft to extract ad headlines.

Step-by-Step: What I Actually Did§

I wrote a coordinator script that manages the state. The user enters a topic. The script sends the topic to the first prompt, which returns a Markdown outline table.

The output is then passed as context to the second prompt. The second prompt expands each outline section into complete paragraphs. Finally, the generated article body is sent to the third model call, which extracts high-CTR Google Ad headlines.

Code Samples / Prompts Used§

Below is the initial prompt in the chain, designed to structure outline fields before content generation begins:

[Loading prompt card for DeepSeek Chat...]

The output schema was strictly validated before triggering the copywriting API call.

What Worked Well§

Splitting the task allowed the models to concentrate on formatting. The outline tables were perfectly aligned, and the final ad copy was optimized for specific keywords. Using JSON format responses made parsing between the script and APIs trivial.

What Failed and Why§

Handling network errors during middle-chain steps was challenging. If prompt 2 timed out, prompt 3 would receive empty variables, triggering a crash. I had to implement a local cache layer using SQLite to store checkpoint outputs. This allowed the script to retry from the failed step instead of restarting from the beginning.

TIP

Optimization Tip: Always set explicit retry parameters and timeout limits of 10-15 seconds for mid-workflow API requests.

Results and Takeaways§

Chaining increased output quality ratings by 35% compared to single-shot prompts. The execution time for the entire workflow was 14 seconds, costing less than $0.01 in tokens. The pipeline is highly scalable and handles content generation smoothly.

Try It Yourself§

Customize variables and test multi-step prompt chains directly inside our live sandbox.