Skip to main content

How Workshop builds

Workshop defaults to Build Mode — you describe what you want, and it starts building. For most tasks, this is all you need. Describe the goal, provide context, and iterate on the results. When a project needs more upfront thinking, Plan Mode is available as a scoping tool. It acts like an experienced technical product manager — asking questions, pushing back on scope, and producing a structured plan before any code gets written. Neither mode is “the right way.” Build Mode is the default because most interactions are iterative. Plan Mode is there when you need it.

Model selector

Workshop is multi-model and multi-provider — you choose which model powers your main agent. Open the model selector in the top-left of the conversation to switch at any time. For most projects, we recommend a strong orchestrator model like Claude Sonnet 4.6, GLM-5.1, or GPT-5.4. These models are especially good at breaking down complex tasks, using tools effectively, and delegating work to sub-agents when needed. If you need maximum performance on a complex project, top-tier models like Claude Opus will produce excellent results — just note they consume more credits. A strong orchestrator will also delegate simpler subtasks to faster, cheaper models automatically, so you’re not paying top-tier rates for everything. Smaller models like Gemini 3.1 Flash or Claude Haiku 4.5 are a good choice for quick, focused tasks that aren’t part of a larger project — reading a document, summarizing something, or answering a question. They’re fast and cost-efficient, but may underperform on multi-step builds or complex orchestration. If you’re using Workshop Desktop, local models are also available in the selector — see Local Models to set them up.
Open the model selector and click “What’s this?” at the top for a detailed breakdown of what each model is best at.
The currently available models are:
ModelProviderBest forSpeedCost
Claude Sonnet 4.6AnthropicBalanced — code, review, orchestrationFastMedium
Claude Haiku 4.5AnthropicFast, focused tasks — searches, scans, checklistsFastestLow
Claude Opus 4AnthropicComplex reasoning — architecture, multi-file analysisSlowerHighest
GPT-5.4OpenAIBackend coding, code review, general-purpose reasoningFastMedium
GPT-5.4 MiniOpenAILighter tasks at lower costFastLow
GPT-5.4 NanoOpenAISimple tasks, lowest costFastestLowest
Gemini 3.1 ProGoogleFrontend dev, long-context analysis (1M token window)FastMedium
Gemini 3.1 FlashGoogleFast general-purpose tasks, large contextFastestLow
GLM-5.1Z.aiFrontend dev, long-horizon tasks, cost-efficient codingFastLowest
GLM-5Z.aiGeneral-purpose codingFastLow
GLM-5 TurboZ.aiFast coding tasksFastestLowest
GLM-5v TurboZ.aiVision + coding tasksFastLow
Kimi K2.5MoonshotCoding, reasoning, multilingual tasksFastLow
This list is updated regularly as new models become available. For the latest, check the model selector in the app.
For more on how different models work together in a single workflow, see Sub-agents.

Communicating effectively

The clearer you are about what you want, the better Workshop can deliver. Think of it as a collaborative conversation.

Be specific and descriptive

Vague prompts produce generic results. Specificity gives Workshop the information it needs to make the right decisions.
Make me a website.

Provide context and purpose

Tell Workshop why you’re building something, not just what. Context about your situation, audience, and constraints leads to better decisions.
Build a database.

Iterate and refine

Building software is iterative. Instead of expecting perfection on the first try, guide Workshop through incremental improvements with specific feedback.
This isn't right. Fix it.

Prompting by intent

Starting a project

Give Workshop a clear picture of what you’re building, who it’s for, and what success looks like.
I'd like to create a fitness tracking app called "FitTracker". The goal is to
monitor workouts, nutrition, and progress toward specific goals. Users should
see a dashboard summarizing their activity, calorie intake, and workout
history. Include data visualization for tracking progress.

Describing design and UI

Reference specific styles, colors, layouts, and interaction patterns rather than abstract adjectives.
The interface should follow a minimalist design with navy, white, and light gray.
Use subtle shadows and rounded corners for a modern feel.
Add a navigation menu with Home, Products, Services, About, and Contact. Fixed
at the top, collapses into a hamburger on mobile. Highlight the active page.

Defining functionality

Describe user interactions as specific scenarios — what the user does, what should happen, and what edge cases to handle.
When a user clicks "Submit", validate all form fields. If validation fails,
show error messages below the problematic fields. On success, send the data
and show a confirmation.

Backend and API work

Specify data models, integrations, performance requirements, and security considerations.
Connect to the OpenWeatherMap REST API. Authenticate with our API key, fetch
current weather and 5-day forecast, and cache responses for 30 minutes.

Debugging and troubleshooting

Describe what you expected, what actually happened, and any relevant context about when the issue occurs.
The contact form fails silently when attachments exceed 5MB. Instead of a blank
error page, show a user-friendly message with the file size limit.

Requesting changes

Be specific about what to change, where to change it, and why.
On the product detail page: 1) Move description above specifications,
2) Make images larger with a zoom feature, 3) Add a "Frequently bought
together" section at the bottom.

Plan Mode

Plan Mode is an optional scoping tool for when you want structured requirements gathering before building. It’s especially useful when starting a new project from scratch or when you’re not sure exactly what you need.

What Plan Mode does

  • Asks probing questions to understand your true needs — typically 2–3 structured questions per turn
  • Scopes down to an MVP before expanding — build something valuable first
  • Pushes back when you’re over-engineering or adding unnecessary complexity
  • Produces a plan document that Build Mode can follow step by step

How it works

1

Discovery

Plan Mode asks questions to understand your needs using structured multiple-choice for clear, efficient communication.
2

MVP Scoping

Based on your answers, Plan Mode identifies the smallest version that delivers real value. It steers away from scope creep.
3

Summary Checkpoint

Before generating the plan, Plan Mode presents a concise summary of scope and technical decisions. You review and approve — or request adjustments.
4

Plan Generation

Plan Mode creates a structured plan.md file saved in your project under ./plans/[timestamp_topic-name]/.

The plan document

Each plan is append-only — new plans never overwrite existing ones, preserving your decision history. The plan includes:
  • Spec Header — Name, smallest scope, non-goals
  • Decision Snapshot — Key technical and product decisions
  • Architecture at a Glance — Recommended stack and structure
  • Implementation Plan — Step-by-step development phases
  • Verification & Demo Script — How to test each phase
  • Deploy — Deployment strategy
Plan Mode is read-only — it cannot modify files or execute code. Planning stays separate from implementation.

Switching between modes

Open the + menu to the left of the conversation input and choose Mode: Build to reveal the mode picker, then select Plan or Build. You can also just type “Use Plan mode.” (or “Switch to Build mode.”) — Workshop will flip the toggle for you. The mode persists across sessions and defaults to Build Mode. Plan Mode menu under the + button, with Build and Plan options A common workflow for new projects:
  1. Plan Mode — Define requirements, scope the MVP, create a plan
  2. Build Mode — Implement Phase 1
  3. Plan Mode — Evaluate results, plan Phase 2
  4. Build Mode — Continue implementation
For existing projects or clear requirements, skip Plan Mode entirely and work in Build Mode.

Voice input

Workshop supports voice dictation — speak your prompt instead of typing it. Your speech is transcribed and inserted into the message input. How to use it:
  • Click the microphone icon in the message input toolbar, or press D (Mac) / CtrlD (Windows/Linux)
  • Speak your prompt — the button pulses while recording
  • Click the microphone again (or press the shortcut again) to stop — Workshop transcribes your audio and inserts the text
  • Edit the transcription if needed, then send as usual
Voice input works in both Workshop Cloud and Desktop. Audio is sent to a transcription service for processing — it is not stored after transcription completes.
Voice input is especially useful for describing UI changes, explaining bugs, or dictating long prompts where typing would be slower.

Files and attachments

You can attach files directly in the message input — but how you attach a file determines what Workshop can do with it. Attaching in the input box (drag-and-drop, paste, or clipboard) The model can read and understand the file — useful for showing Workshop a design reference, a screenshot of a bug, or a spec to follow. The file is not stored anywhere your app can access, so it won’t appear in what gets built. Uploading a file for your app to use If you want Workshop to use a file inside your app — a logo, a background image, a data file — you need to upload it first:
  1. Click the + button in the message input and choose Files & Uploads, or go to Workshop Hub → Uploads
  2. Type @ in the message input and select the file by name
  3. Tell Workshop how to use it — e.g. “Use @logo.png as the logo in the header”
Context vs content — when it matters:
  • Context files help Workshop understand your project — reference material, design specs, data schemas, or documentation it should read but not directly include in output. Add files to your project’s context through the sidebar Sources panel.
  • Content files are material Workshop should use directly — images to display, data files to process, CSVs to transform. Upload these and reference with @ in your message.
For persistent context across conversations, use the Sources panel in the sidebar or the .workshop/ directory structure described in Context and Memory.

Message queuing

You can keep typing while Workshop is responding. Messages you send during an active response are queued and delivered automatically once the current response completes. This means you don’t need to wait for the agent to finish before giving your next instruction — just type and send as you think of things. Queued messages appear above the input so you can review or remove them before they’re sent.

Send now

If you want your queued messages sent immediately without waiting for the current response to finish, click Send now. This interrupts the agent’s current response and delivers your queued messages as a single combined prompt.

Managing long conversations

Every conversation has a finite context window — the amount of information Workshop can hold in memory at once. As a conversation grows, responses can slow down and quality can degrade toward the end of the window. Workshop handles this automatically so you can focus on building.

Auto-compaction

When your conversation approaches the context window limit, Workshop automatically compacts it. This creates an intelligent summary of the conversation so far and continues seamlessly — no action needed from you. Here’s what happens:
  1. Workshop detects that the conversation is nearing its context limit
  2. It generates a focused summary that preserves the key decisions, code context, and current task state
  3. Your conversation continues in a new child conversation, carrying forward the essential context
  4. The original conversation becomes read-only history you can review anytime
Auto-compaction is always active. You don’t need to configure or enable anything — long conversations stay fast without manual intervention.

The /compact command

You can also compact manually at any time using the /compact command. This is useful when you want to start a fresh context on your own terms — for example, after completing a milestone. When a conversation nears the context limit, a banner appears suggesting compaction: Workshop suggests compaction when context is running low You can click Compact conversation and continue in the banner, or type /compact in the chat input at any time.
/compact
Summarizes everything and continues in a new conversation.
/compact focus on the authentication flow
Prioritizes specific context in the summary. Workshop analyzes the full conversation and distills it into a structured summary — capturing your original request, key decisions, current project state, and what to do next. Once complete, the conversation is replaced with a clean summary. You can expand it to review what was captured, and continue prompting as usual — the agent retains full understanding of what was built. The conversation after compaction — a clean summary with full project context preserved

Other useful commands

CommandDescriptionExample
/context [guidance]Update project context from conversation/context focus on API patterns
/summarize [focus]Summarize the current conversation/summarize action items

Tips for better results

Use version control

Projects can evolve quickly and it’s easy to lose a working state. Ask Workshop to initialize Git and commit at key milestones so you can always go back to a version that worked — or connect a GitHub repository using the native GitHub integration to back up your code and collaborate with others.

Try Plan Mode for new projects

When starting from scratch, Plan Mode helps you define clear requirements and creates a structured plan — even from vague ideas.

Work incrementally

Build features one at a time. Get core functionality working before adding complexity. This makes debugging easier and progress more visible.

Use custom instructions

Set consistent preferences in Settings so you don’t repeat the same guidance in every conversation.

Leverage project context

Keep your .workshop/context.md updated so Workshop always understands your project’s current state and conventions. See Context and Memory.