> For the complete documentation index, see [llms.txt](https://manual.formwise.ai/formwise/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://manual.formwise.ai/formwise/advanced-features/prompt-engineering/advanced-prompt-chaining-guide.md).

# Advanced Prompt Chaining Guide

Advanced Best Practices for Prompt Chaining in SmartForms

SmartForms support **prompt chaining** for multi-step AI workflows. Each step builds on earlier results. This improves structure and consistency.

Example workflow:

* Generate a customer avatar.
* Draft a value proposition.
* Convert it into an email sequence.

***

### How prompt chaining works

When a SmartForm has **Prompt Instructions**, it runs one model call. That call uses mapped form fields and any backend data sources.

You can add more prompt steps. Each step runs as a **separate API call**. Each step produces its own output.

Later steps can:

* Reuse earlier outputs with `@step_1`, `@step_2`, and so on.
* Run a new task with different instructions and context.

***

### Reference earlier steps with `@step_x`

Use `@step_1`, `@step_2`, and so on inside later instructions. Treat each reference as raw text from the earlier output.

{% code title="Example" %}

```plaintext
Write a cold email introducing the new offer.
Use the customer avatar below.

Customer avatar:
@step_1
```

{% endcode %}

Best practices:

* Add a label before the reference.
* State what the referenced content represents.
* Specify the output format you want.

Avoid vague instructions like:

```plaintext
Use @step_1 to generate the copy.
```

Prefer explicit instructions like:

```plaintext
Using the brand voice below, write 5 Google ad headlines.

Brand voice:
@step_1
```

***

### Prompt chaining design patterns

#### Define clear stages

Split the workflow into small steps. Keep each step focused.

Common staging:

* Step 1: Generate core data (avatar, outline, research summary).
* Step 2: Generate the primary asset (copy, plan, script).
* Step 3: Refine and format (tone, structure, length).
* Step 4: Repurpose (emails, ads, bullets, slides).

#### Make each step self-contained

Do not assume the model remembers your intent. Restate the goal of the step. Explain what the referenced content contains.

Example:

```plaintext
Rewrite the value proposition in a more emotional tone.
Use the tone guide below.

Tone guide:
@step_1
```

#### Reuse SmartForm fields in later steps

You can mix `@step_x` outputs with mapped fields like `@q1answer`.

```plaintext
Write 3 Instagram captions for @q1answer.
Use the product description and brand voice below.

Product description:
@step_1

Brand voice:
@step_2
```

***

### Token, cost, and timeout considerations

{% hint style="warning" %}
Each prompt step is a separate API call. Each call has its own token budget. Large outputs early can reduce quality later. They also increase cost and latency.
{% endhint %}

Practical guidance:

* Use GPT-5.2 for deeper reasoning and better formatting.
* Keep step outputs short when they will be reused downstream.
* Avoid repeating large backend content in every step.
* Avoid stacking web browsing, scraping, and large multi-step chains.

Common symptoms of oversized chains:

* Truncated outputs.
* A step failing mid-run.
* The full SmartForm timing out.

***

### Data source considerations

Backend data sources can be expensive in tokens. Referencing them in many steps increases usage.

Recommended pattern:

* Inject the data source in step 1.
* Summarize or extract what you need.
* Pass the summary downstream via `@step_1`.

***

### Checklist

* Break the workflow into clear steps.
* Add context labels before `@step_x` references.
* Keep reusable step outputs concise.
* Pick a model that matches your token needs.
* Avoid re-ingesting large data sources across steps.
* Test each step before publishing the SmartForm.