Workflow Design Best Practices for PromptPlan

Keep each step single-purpose

Good workflow steps do one thing clearly:

• normalize input
• plan
• generate
• review
• export

Bad workflow steps try to analyze, draft, critique, and format all at once.

Name inputs and steps like stable API fields

Use names you will still understand later:

• product_context
• feature_description
• review_edge_cases
• export_plan

Stable names make prompt references safer and clearer.

Put stable logic in prompts, changing context in inputs

Do not turn every instruction into a workflow input.

Use inputs for values that change often:

• audience
• source material
• constraints
• target platform

Keep the durable reasoning structure in the prompt template.

Separate generation from review

If quality matters, use at least two steps:

1. generate the draft
2. review or improve the draft

This makes failures easier to diagnose and reruns much cheaper.

Add an explicit export step

Use a final export step when the workflow output needs a consistent shape.

That step is the best place to:

• standardize headings
• package earlier outputs together
• remove repetition
• prepare markdown or JSON-friendly output

Start cheap, pin only where needed

Do not pin models everywhere by default.

A good starting pattern is:

• let run mode handle defaults
• pin a stronger model only on reasoning or critical generation steps
• keep export and formatting steps cheap

Design for reruns

Assume one step will need adjustment later.

That means:

• avoid giant prompts
• keep step outputs reusable
• avoid hidden assumptions between far-apart steps
• use named step references when a later step depends on a specific artifact

Use starter workflows as scaffolds

Starter workflows are a good base, not a fixed endpoint.

Use them to:

• learn a good workflow shape
• shorten setup time
• fork or duplicate into your own version
• replace generic prompts with domain-specific instructions