Which models are available per step?

Sixteen across seven providers: OpenAI (GPT-4o, GPT-4o mini, GPT-4 Turbo, o1, o3-mini), Anthropic (Claude 3.5 Sonnet, 3.5 Haiku, 3 Opus), Google (Gemini 1.5 Pro, 2.0 Flash), Meta (Llama 3.1 70B and 405B), Mistral (Large, Mixtral 8x7B), Cohere (Command R+), and DeepSeek (V3). New steps default to GPT-4o; the model select shows each entry with its provider in parentheses.

What does the temperature slider control?

The per-step randomness setting, ranging from 0 (deterministic and focused) to 2 (highly creative) in 0.05 increments, defaulting to 0.7 for new steps. The value is stored in the exported JSON and shown as T=0.70 on the flow diagram. The Notes card suggests 0.0–0.7 for classification-style steps and 0.7–1.0 for generation.

How are {{variables}} handled in steps?

Each step's prompt template is scanned for {{name}} placeholders with the same regex used elsewhere on the site. Detected variables render as badges under the step and accumulate in the Variables across workflow card — the combined list of runtime inputs your pipeline must supply. They are placeholders only: no values are substituted or executed in the designer.

What format does Export JSON produce?

A versioned payload: { version: 1, exportedAt: ISO timestamp, steps: [...] }, where each step carries id, name, prompt, model, and temperature. The file downloads as ai-workflow.json, and Copy JSON puts the same content on the clipboard. Import accepts both this wrapped shape and a bare array of steps, replacing the current design.

Where is my workflow design saved?

Automatically in localStorage under fl-ai-workflow, so reloading the page restores your steps. Like all localStorage data it is browser- and device-specific, and a site-data wipe removes it. Export JSON regularly — the versioned file is both your backup and the artifact your execution pipeline will consume.

Misc

AI Workflow Builder

Design multi-step AI workflows with a visual flow diagram.

aiworkflowpipelinestepsdesign

Loading tool…

About the AI Workflow Builder

The AI Workflow Builder is a visual designer for multi-step LLM pipelines. Every step is a card holding a name, a prompt template with {{variable}} placeholders, a model selected from sixteen options across OpenAI, Anthropic, Google, Meta, Mistral, Cohere, and DeepSeek, and a temperature slider from 0 to 2. Steps run top to bottom, and a live flow diagram renders each as a numbered node with its model and T= badge, joined by arrows so the execution order is unmistakable. A Variables across workflow card aggregates every placeholder detected in the pipeline into a single checklist of runtime inputs. The tool is explicit that it designs rather than executes: no LLM API is ever called, and the finished artifact is a versioned JSON payload with an exportedAt timestamp that you can copy or download and feed into your own runner. Edits persist automatically in localStorage, and import/export lets the design travel between machines. For anyone planning a multi-step AI workflow — extraction, classification, generation — it turns a vague idea into a concrete, reviewable spec.

Hand-written guide

Examples

Input

Click Add step twice. Step 1: rename to "Extract entities", prompt "Extract named entities from {{article}} and return them as JSON.", model GPT-4o mini, temperature 0.10. Step 2: rename to "Summarise", prompt "Summarise the entity list into a 3-bullet brief for {{audience}}.", keep GPT-4o and the 0.7 default.

Output

Stats: Steps 2 · Distinct models 2 · Variables 2. Each step card lists its own variables (article on step 1; audience on step 2), the Variables across workflow card shows {{article}} and {{audience}} once each, and the Flow diagram draws two numbered nodes joined by an arrow, with GPT-4o mini / T=0.10 and GPT-4o / T=0.70 badges.

Note: New steps arrive named "Step 1", "Step 2"… with GPT-4o and temperature 0.7 preselected.

Input

With those two steps in place, click Export JSON.

Output

{
  "version": 1,
  "exportedAt": "<ISO timestamp>",
  "steps": [
    {
      "id": "s-<generated>",
      "name": "Extract entities",
      "prompt": "Extract named entities from {{article}} and return them as JSON.",
      "model": "gpt-4o-mini",
      "temperature": 0.1
    },
    {
      "id": "s-<generated>",
      "name": "Summarise",
      "prompt": "Summarise the entity list into a 3-bullet brief for {{audience}}.",
      "model": "gpt-4o",
      "temperature": 0.7
    }
  ]
}

Note: Downloads as ai-workflow.json. Copy JSON writes the identical payload to the clipboard.

Input

Import a JSON file — either this exported shape or a bare array of steps.

Output

Valid steps (a prompt string and a model string are required) replace the current design and a toast reports "Imported N step(s)". Missing ids are regenerated and a missing temperature falls back to 0.7.

Note: Import replaces rather than merges — export first if you want to keep the current design.

How to use

1
Click Add step to create the first step card — it arrives named "Step 1" with the GPT-4o model and a 0.7 temperature.
2
Edit the step: rename it in the Step name field and write the prompt template in the Prompt template box, using {{variables}} for runtime inputs.
3
Choose a Model from the sixteen-provider dropdown and drag the Temperature slider (0–2) to match the step's purpose.
4
Reorder steps with the up/down arrows or remove them with the trash icon; the Flow diagram below redraws the sequence with arrows.
5
Check the Steps / Distinct models / Variables stats and the Variables across workflow card, then Export JSON (or Copy JSON) for your pipeline.

Common use cases

Design a content pipeline — extract keywords, generate an outline, then draft the article, each as its own step.
Plan a multi-model pipeline — cheap models (GPT-4o mini, Haiku) for extraction steps and expensive ones (GPT-4o, Opus) for synthesis.
Define a data-analysis chain — classify, extract, and summarise in sequence with low temperatures for consistency.
Sketch a support bot — intent detection, response drafting, and tone-check steps mapped visually before any code is written.
Document a workflow for engineering — the exported JSON with version and timestamps doubles as a spec for implementation.
Teach pipeline design — the flow diagram makes the concept of step order and variable flow concrete for learners.

Best practices

Set temperature per step by purpose: use 0.0–0.3 for extraction and classification steps and 0.7–1.0 for generation — the slider's hint marks 0 as focused and 2 as creative.
Keep model choice per step intentional: the designer tracks "Distinct models", so a pipeline that uses GPT-4o mini for extraction and GPT-4o for synthesis is a feature, not an accident.
Use {{variables}} in every step template: the Variables overview card lists placeholders across the whole workflow, giving you a single checklist of runtime inputs before you wire the pipeline.
Remember the tool designs but does not execute: no API calls are made, so prompts can reference any model in the list safely, but you must build your own runner from the exported JSON.
Reorder with the arrows before exporting: step order in the JSON is execution order, and the flow diagram is your visual proof that the sequence reads top-to-bottom.
Export after every design session: the payload includes a version field and exportedAt timestamp, which makes each ai-workflow.json a dated snapshot you can diff in source control.

Tips

The flow diagram reuses the live step data, so rename a step and the diagram updates instantly — keep names short and action-oriented.
Temperature is per step in the exported JSON, so a mixed pipeline can stay deterministic upstream and creative downstream.
Import replaces the current design rather than merging — export first if the existing steps matter.
Keep {{variable}} names consistent across steps: the Variables overview deduplicates, so the same placeholder appears once no matter how many steps use it.

Frequently asked questions

No. It is explicitly a designer: the alert at the top states "Designer only — no execution", and no LLM API is called at any point. You define steps, models, and temperatures, then export the JSON and wire it into your own runner — the exported payload contains everything a pipeline needs except the code that runs it.

Explore more misc tools

Browse the full collection of misc tools on the hub, or jump back to all categories.

All Misc tools All categories