Introducing DeepWork: structured workflow automation.
We're open-sourcing an internal tool that adds process discipline to AI-assisted tasks. It's early, but it works.
January 2025
Why we built it
Today we're releasing DeepWork, a workflow automation framework built on top of Claude Code. It's available now on GitHub.
We kept running into a frustrating pattern with AI tools. They're powerful, sometimes shockingly so, but the quality of results varies wildly depending on how you prompt them. Ask the same question five different ways, get five different answers. Forget to include context, get an incomplete result. Skip a verification step, get something that looks right but isn't.
The problem isn't the AI. It's that complex tasks need structure. Data analysis isn't just “analyze this data.” It's: connect to sources, understand the schema, define clear questions, ground your analysis in sanity checks, look for gaps, validate findings, and then present results. Skip steps, and you get garbage.
DeepWork is our attempt to encode that structure. Instead of relying on prompts and hoping the AI follows a good process, we define workflows with explicit steps, dependencies between steps, and quality gates that verify outputs before proceeding.
How it works
Define tasks in plain English. Describe what you want, and DeepWork breaks it into steps.
Quality gates at each step. Stop hooks verify outputs before the workflow proceeds.
Everything is versioned. All work products and intermediate steps live in Git.
Who it's for
DeepWork is designed for people who want to automate complex, multi-step tasks but don't want to write code. Analysts, operators, founders: anyone who needs to synthesize research, analyze data, write content, or run outreach at scale.
This isn't just aspirational. We've validated it with users who had never opened a terminal before their first DeepWork session. They successfully completed real tasks: data consolidation, report generation, research synthesis. The learning curve is there, but it's manageable.
What's included
DeepWork ships with pre-built workflows for common tasks. Use them as-is, or as templates for building your own.
Data Analysis
Connect to sources, define questions, analyze, validate, present.
Blog Post
Research, outline, draft, review, edit, SEO optimization.
Competitive Research
Systematic discovery and analysis of competitors.
Founder Outreach
Research targets, personalize messages, track responses.
Still figuring out
DeepWork is early. We're releasing it now because it's useful, not because it's polished. Some known issues:
Setup is fiddly. You need Python 3.11, Claude Code installed, and some terminal comfort. We're working on simplifying this.
Sometimes the AI escapes. It occasionally decides to skip the structured workflow and "just do the thing." We're improving the guardrails.
Documentation is sparse. We prioritized shipping. Better docs are coming.
If you run into problems or have ideas for improvements, open an issue on GitHub. This is a work in progress, and feedback shapes what we build next.
Getting started
terminal
$ pip install deepwork$ deepwork run data_analysis
Full setup instructions are in the GitHub README. For a deeper look at what DeepWork can do, see the landing page.