Workflows

Workflows are like prompts on steroids. They harness the untapped power and control of LLMs through progressive disclosure—breaking complex tasks into focused steps that execute sequentially. Instead of random AI slop where you hope for the best, workflows give you repeatable, reliable, high-quality outputs.

This guide explains what workflows are, why they’re powerful, and how to think about designing them.

What Is a Workflow?

A workflow is a structured process where the AI executes steps sequentially to accomplish a task. Each step has a specific purpose, and the AI moves through them methodically—whether that involves extensive collaboration or minimal user interaction.

Think of it this way: instead of asking “help me build a nutrition plan” and getting a generic response, a workflow guides you (or runs automatically) through discovery, assessment, strategy, shopping lists, and prep schedules—each step building on the last, nothing missed, no shortcuts taken.

How do workflows differ from skills?

Actually they really do not - a workflow can be a skill, and a skill can be a workflow. The main thing with a BMad workflow is the suggestion to follow certain conventions, which actually are also skill best practices. A skill has a few optional and required fields to add as the main file workflow and get stored in a specific location depending on your tool choice for automatic invocation by the llm - whereas workflows are generally intentionally launched, with from another process calling them, or a user invoking via a slash command. In the near future, workflows will optionally be installable as skills also - but if you like, you can add front matter to your custom workflows based on the skill spec from Anthropic, and put them in the proper location your tool dictates.

The Power of Progressive Disclosure

Here’s why workflows work so well: the AI only sees the current step. It doesn’t know about step 5 when it’s on step 2. It can’t get ahead of itself, skip steps, or lose focus. Each step gets the AI’s full attention, completing fully before the next step loads.

This is the opposite of a giant prompt that tries to handle everything at once and inevitably misses details or loses coherence.

Workflows exist on a spectrum:

Interactive workflows guide users through complex decisions via collaboration and facilitation
Automated workflows run with minimal user input, processing documents or executing tasks
Hybrid workflows combine both—some steps need user input, others run automatically

Real-World Workflow Examples

Tax Organizer Workflow

A tax preparation workflow that helps users organize financial documents for tax filing. Runs in a single session, follows prescriptive IRS categories, produces a checklist of required documents with missing-item alerts. Sequential and compliance-focused.

Meal Planning Workflow

Creates personalized weekly meal plans through collaborative nutrition planning. Users can stop mid-session and return later because the workflow tracks progress. Intent-based conversation helps discover preferences rather than following a script. Multi-session, creative, and highly interactive.

Course Creator Workflow

Helps instructors design course syllabi. Branches based on course type—academic courses need accreditation sections, vocational courses need certification prep, self-paced courses need different structures entirely.

Therapy Intake Workflow

Guides mental health professionals through structured client intake sessions. Highly sensitive and confidential, uses intent-based questioning to build rapport while ensuring all required clinical information is collected. Continuable across multiple sessions.

Software Architecture Workflow (BMM Module)

Part of a larger software development pipeline. Runs after product requirements and UX design are complete, takes those documents as input, then collaboratively walks through technical decisions: system components, data flows, technology choices, architectural patterns. Produces an architecture document that implementation teams use to build consistently.

Shard Document Workflow

Nearly hands-off automated workflow. Takes a large document as input, uses a custom npx tool to split it into smaller files, deletes the original, then augments an index with content details so the LLM can efficiently find and reference specific sections later. Minimal user interaction—just specify the input document.

These examples show the range: from collaborative creative processes to automated batch jobs, workflows ensure completeness and consistency whether the work involves deep collaboration or minimal human oversight.

The Facilitative Philosophy

When workflows involve users, they should be facilitative, not directive. The AI treats users as partners and domain experts, not as passive recipients of generated content.

Collaborative dialogue, not command-response: The AI and user work together throughout. The AI brings structured thinking, methodology, and technical knowledge. The user brings domain expertise, context, and judgment. Together they produce something better than either could alone.

The user is the expert in their domain: A nutrition planning workflow doesn’t dictate meal plans—it guides users through discovering what works for their lifestyle. An architecture workflow doesn’t tell architects what to build—it facilitates systematic decision-making so choices are explicit and consistent.

Intent-based facilitation: Workflows should describe goals and approaches, not scripts. Instead of “Ask: What is your age? Then ask: What is your goal weight?” use “Guide the user through understanding their health profile. Ask 1-2 questions at a time. Think about their responses before asking follow-ups. Probe to understand their actual needs.”

The AI figures out exact wording and question order based on conversation context. This makes interactions feel natural and responsive rather than robotic and interrogative.

When to be prescriptive: Some workflows require exact scripts—medical intake, legal compliance, safety-critical procedures. But these are the exception, not the rule. Default to facilitative intent-based approaches unless compliance or regulation demands otherwise.

Why Workflows Matter

Workflows solve three fundamental problems with AI interactions:

Focus: Each step contains only instructions for that phase. The AI sees one step at a time, preventing it from getting ahead of itself or losing focus.

Continuity: Workflows can span multiple sessions. Stop mid-workflow and return later without losing progress—something free-form prompts can’t do.

Quality: Sequential enforcement prevents shortcuts. The AI must complete each step fully before moving on, ensuring thorough, complete outputs instead of rushed, half-baked results.

How Workflows Work

The Basic Structure

Workflows consist of multiple markdown files, each representing one step:

my-workflow/
├── workflow.md              # Entry point and configuration
├── steps/                   # Step files (steps-c/ for create, steps-e/ for edit, steps-v/ for validate)
│   ├── step-01-init.md
│   ├── step-02-profile.md
│   └── step-N-final.md
├── data/                    # Reference materials, CSVs, examples
└── templates/               # Output document templates

The workflow.md file is minimal—it contains the workflow name, description, goal, the AI’s role, and how to start. Importantly, it does not list all steps or detail what each does. This is progressive disclosure in action.

Sequential Execution

Workflows execute in strict sequence: step-01 → step-02 → step-03 → ... → step-N

The AI cannot skip steps or optimize the sequence. It must complete each step fully before loading the next. This ensures thoroughness and prevents shortcuts that compromise quality.

Continuable Workflows

Some workflows are complex enough that users might need multiple sessions. These “continuable workflows” track which steps are complete in the output document’s frontmatter, so users can stop and resume later without losing progress.

Use continuable workflows when:

The workflow produces large documents
Multiple sessions are likely
Complex decisions benefit from reflection
The workflow has many steps (8+)

Keep it simple (single-session) when tasks are quick, focused, and can be completed in one sitting.

Workflow Chaining

Workflows can be chained together where outputs become inputs. The BMM module pipeline is a perfect example:

brainstorming → research → brief → PRD → UX → architecture → epics → sprint-planning
                                                        ↓
                                            implement-story → review → repeat

Each workflow checks for required inputs from prior workflows, validates they’re complete, and produces output for the next workflow. This creates powerful end-to-end pipelines for complex processes.

For critical workflows that produce important artifacts, BMad uses a tri-modal structure: Create, Validate, and Edit. Each mode is a separate workflow path that can run independently or flow into the others.

Create mode builds new artifacts from scratch. But here’s where it gets interesting: create mode can also function as a conversion tool. Feed it a non-compliant document—something that doesn’t follow BMad standards—and it will extract the essential content and rebuild it as a compliant artifact. This means you can bring in existing work and automatically upgrade it to follow proper patterns.

Validate mode runs standalone and checks artifacts against standards. Because it’s separate, you can run validation whenever you want—immediately after creation, weeks later when things have changed, or even using a different LLM entirely. It’s like having a quality assurance checkpoint that’s always available but never forced.

Edit mode modifies existing artifacts while enforcing standards. As you update documents to reflect changing requirements or new understanding, edit mode ensures you don’t accidentally drift away from the patterns that make the artifacts useful. It checks compliance as you work and can route back to create mode if it detects something that needs full conversion.

All BMad planning workflows and the BMB module (will) use this tri-modal pattern. The pristine example is the workflow workflow in BMB—it creates workflow specifications, validates them against standards, and lets you edit them while maintaining compliance. You can study that workflow to see the pattern in action.

This tri-modal approach gives you the best of both worlds: the creativity and flexibility to build what you need, the quality assurance of validation that can run anytime, and the ability to iterate while staying true to standards that make the artifacts valuable across sessions and team members.

Design Decisions

Before building a workflow, answer these questions:

Module affiliation: Is this standalone or part of a module? Module-based workflows can access module-specific variables and reference other workflow outputs. Also when part of a module, generally they will be associated to an agent.

Continuable or single-session?: Will users need multiple sessions, or can this be completed in one sitting?

Edit/Validate support?: Do you need Create/Edit/Validate modes (tri-modal structure)? Use tri-modal for complex, critical workflows requiring quality assurance. Use create-only for simple, one-off workflows.

Document output?: Does this produce a persistent file, or perform actions without output?

Intent or prescriptive?: Is this intent-based facilitation (most workflows) or prescriptive compliance (medical, legal, regulated)?

Learning from Examples

The best way to understand workflows is to study real examples. Look at the official BMad modules:

BMB (Module Builder): Workflow and agent creation workflows
BMM (Business Method Module): Complete software development pipeline from brainstorming through sprint planning
BMGD (Game Development Module): Game design briefs, narratives, architecture
CIS (Creativity, Innovation, Strategy): Brainstorming, design thinking, storytelling, innovation strategy

Study the workflow.md files to understand how each workflow starts. Examine step files to see how instructions are structured. Notice the frontmatter variables, menu handling, and how steps chain together.

Copy patterns that work. Adapt them to your domain. The structure is consistent across all workflows—the content and steps change, but the architecture stays the same.

When to Use Workflows

Use workflows when:

Tasks are multi-step and complex: Break down complexity into manageable pieces
Quality and completeness matter: Sequential enforcement ensures nothing gets missed
Repeatability is important: Get consistent results every time
Tasks span multiple sessions: Continuable workflows preserve progress
You need to chain processes: Output of one workflow becomes input of another
Compliance or standards matter: Enforce required steps and documentation

Don’t use workflows when:

Tasks are simple and one-off: A single prompt works fine for quick questions
Flexibility trumps structure: Free-form conversation is better for exploration

Modified BMad Workflows

Tasks are truly one-step

If there’s only one thing to do and it can be explained in under about 300 lines - don’t bother with step files. Instead, you can still have a short single file workflow.md file.

The Bottom Line

Workflows transform AI from a tool that gives variable, unpredictable results into a reliable system for complex, multi-step processes. Through progressive disclosure, sequential execution, guided facilitation, and thoughtful design, workflows give you control and repeatability that ad-hoc prompting alone can’t match.

They’re not just for software development. You can create workflows for any guided process - meal planning, course design, therapy intake, tax preparation, document processing, creative writing, event planning—any complex task that benefits from structure and thoroughness.

Start simple. Study examples. Build workflows for your own domain. You’ll wonder how you ever got by with just prompts.