What Is the AI Workflow Builder?

The AI Workflow Builder is a visual planner for multi-step AI processes. Instead of describing a pipeline in a document or a ticket, you build it on screen: a list of named steps, each with a prompt template, a model, a temperature and links to the steps that follow it. A flow diagram draws itself as you work, so branches and merge points are visible instead of implied.

The important word is planner. This tool never calls a large language model. There is no API key, no billing and no execution. It resolves your prompt templates against sample variable values, lints the design for structural mistakes, estimates token counts, and exports the result as JSON or Markdown that you can hand to an engineer or feed into your own pipeline.

Workflows are not always linear. A classifier step might route to one of three follow-up prompts, and those branches might merge into a single summarizer step. The builder supports that shape: multiple successors create branches, steps with multiple incoming edges become merge points, cycles are detected and flagged, and the diagram draws arrows, branch arms and merge points automatically.

Why Use It — and Who It Is For

Prompt chains fail in predictable ways: a variable name typed two different ways, a step that references an output nobody produces, a loop nobody noticed, a model chosen for one step that cannot fit the context of the previous step. Catching those on paper or in a code review is slow. Catching them in a live diagram with a linter is fast.

• Product and prompt engineers designing an agent, a classification pipeline or a document processing chain.
• Technical founders sketching a workflow before hiring someone to implement it.
• Consultants who need a readable artifact to align with a client before building.
• Students and researchers structuring an experiment with several prompting stages.
• Teams that want one shared JSON file describing a workflow instead of scattered code comments.

It is also a thinking tool. Many people find that drawing the branches forces decisions that prose hides, such as what happens when the classifier is uncertain or where a human review step belongs.

How This Tool Works

Everything runs client side in your browser. The only network request is the model catalog: the page fetches the public models.dev model list, caches it in local storage for 24 hours, and shows a badge telling you whether the catalog is live or came from the bundled offline list. If the feed is unreachable, a small built-in fallback list keeps the picker usable.

Your workflow itself is autosaved to local storage under the key awb-workflow as you type, and a saved indicator confirms it. It is never uploaded anywhere.

Under the hood the builder maintains a graph:

1. Steps hold a unique name, a prompt template, an optional model, a temperature and a list of next step ids.
2. Edges are derived from the next step selections. A step with more than one outgoing edge is a branch; a step with more than one incoming edge is a merge point.
3. Variables are detected in every prompt with a regular expression that finds placeholder braces. Step output references of the form stepName.output are collected separately and listed as output references.
4. Analysis computes reachable order, branches, paths and cycles for the diagram and for validation.
5. Preview run walks the steps in graph order, replaces each placeholder with the value you supplied, carries forward earlier step outputs, and reports the linter findings next to each resolved prompt.
6. Token estimates use a simple rule of four characters per token, applied per step and summed into a total for the whole workflow.

The linter is deliberately blunt. It reports duplicate step names, steps with no name, empty prompts, missing model selections, unbalanced placeholder braces, empty placeholders, references to unknown step outputs, steps pointing at a missing step, and any cycle in the graph. Fixing these before export is the whole point of the tool.

Key Features

• Branching graph, not just a chain - multiple next steps per step, merge points, and automatic cycle detection.
• Live SVG flow diagram that redraws as you edit; click a node to jump straight to its editor, and cycles are highlighted.
• Step cards with name, reorder up and down, duplicate, remove, prompt template, model picker and temperature slider.
• Next steps chips for routing each step to one or more successors, with a terminal step message when a step ends the workflow.
• Model picker with search, provider grouping and featured chips, plus an offline fallback list when the catalog is unreachable.
• Temperature slider from 0 to 2 that disables itself for models with fixed sampling.
• Variables panel listing every placeholder found across the workflow, with inputs for preview values and a list of step output references.
• Preview run that resolves every prompt with your values and shows per-step token estimates. No model is called.
• Linter covering duplicate names, missing names, empty prompts, missing models, unbalanced or empty placeholders, unknown output references, missing step targets and cycles.
• Three startup templates - linear chain, classifier branch and summarizer chain - to start from a working shape.
• Undo and redo, including the Ctrl+Z shortcut.
• Versioned JSON export in the shape version, exportedAt, title, steps, plus a readable Markdown export for documents.
• JSON import accepting either a bare array of steps or the wrapped export object.
• Statistics bar showing steps, branches, models, variables, estimated tokens, path count and save status at a glance.
• Autosave to browser local storage, with a clear command behind a confirmation prompt.

How to Use — Step by Step

1. Press Add step to create your first step, or pick a starting point from the Templates dropdown: linear chain, classifier branch or summarizer chain.
2. Give the step a unique name. Names matter because step output references use them.
3. Write the prompt template. Any placeholder you type becomes an input variable that you can fill in later for the preview.
4. Press the model button to open the picker, search or browse by provider, and choose a model. Use the clear button if you want the step to stay model agnostic for now.
5. Set the temperature if the selected model supports sampling. The slider disables for models that do not.
6. Use Next steps to route the step. Pick one successor for a chain, or several to create a branch. Leave it empty to mark a terminal step.
7. Repeat for every step. Reorder with the arrows, duplicate a step when the prompt is nearly identical, or remove one with the trash button.
8. Fill the Variables across workflow panel once with sample values for every placeholder you used.
9. Press Preview run. Read the resolved prompts, the per-step token estimates and the linter list. Fix anything marked as an error or warning.
10. Click nodes in the flow diagram to jump to the step that needs attention, and check the statistics bar for branch, path and token totals.
11. When the design is clean, press Export JSON or Copy JSON for your implementation pipeline, or Copy Markdown for a document. Import brings an exported file back later.
12. Use Undo and Redo while experimenting, and Clear when you want to start over. The confirm dialog protects you from losing work by accident.

Limits and Rules

• No execution. The tool never calls a model, needs no API key and bills nothing. Output must be run by your own pipeline.
• No server storage. Your workflow lives in local storage under the key awb-workflow until you clear it or clear browser data. Export a JSON copy for anything important.
• Model catalog comes from models.dev, cached in your browser for 24 hours, with an offline fallback list if the feed cannot be reached. The catalog badge tells you which source is active.
• Token counts are estimates based on four characters per token. They are for sizing a design, not for billing.
• Import needs a usable array of steps. Files that are not valid JSON, that do not contain an array of steps, or that hold no step objects at all are rejected with a clear message. Entries that arrive without a name or without a prompt are imported empty and then flagged by the linter rather than refused.
• Step names must be unique, placeholders must be balanced, and the graph must not contain a cycle, or the linter will flag the problem before you export.
• One workflow at a time. The clear command replaces everything, so export first if you want to keep a copy.
• Use the tool for design and documentation of your own workflows. Do not paste secrets, private keys or credentials into prompts, because the file you export may be shared with your team.

When to Use It — and When Not To

Use it when a workflow has more than one step, when branches or conditionals are involved, when several people need to agree on the shape of a pipeline, or when you want to estimate prompt size and variable requirements before implementation. It is ideal for the design phase, for writing a specification, and for teaching how a chain of prompts behaves.

Do not use it when you need to actually run prompts. There is no execution here, no streaming, no tool calling and no results. If you want to test model outputs, use an API testing tool or your own code with a real key. Likewise, if your workflow is a single prompt with no branching, a plain text editor is faster than a diagram.

Frequently Asked Questions