DeepWork

Structured workflows for long-running AI agents

Describe a process in plain English. DeepWork turns it into a reusable skill with quality gates, so agents stay on-script across complex, high-discretion tasks. Works as a plugin for Claude Code, Codex, and other agent harnesses.

terminal — deepwork
$/deepwork competitive_research
Asking 10 clarifying questions...
Generating workflow: 8 steps, 5 quality gates
✓ Created skill: /competitive_research
$/competitive_research
Scanning company website and social channels...
Identifying competitors from existing list + web search...
Deep research on 6 competitors...
✓ Quality check: SWOT analysis validated
✓ Complete — 4 reports generated on work branch

What is DeepWork?

DeepWork is an open-source plugin for AI agent harnesses — tools like Claude Code, Codex, and Gemini CLI that let AI models operate autonomously on your computer.

These agents are powerful, but they struggle with long, multi-step tasks. They go off-script, lose context between sessions, and skip steps when tasks get complex. Research shows agent failure rates quadruple when you double task duration.

DeepWork solves this by adding workflow structure. Describe a process once, and DeepWork creates a reusable skill — a series of steps with quality gates that keep the agent on track. Skills are composable: you can call one skill from inside another. All work happens on git branches, so everything is versioned and reviewable.

Who it's for

If you can describe a process, you can automate it. DeepWork works for anyone using an AI agent harness.

Founders & operators

Competitive research, investor updates, reports that pull from multiple sources — automated and repeatable.

Product & ops

Tutorial writing, QA reports, daily briefs, SEO analysis, updating process docs — work that recurs but doesn't warrant custom software.

Data & engineering

ETL workflows, custom reports, standup summaries, git log analysis — structured processes with quality checks.

How it works

Quality gates at every step

Stop hooks verify outputs before proceeding. Critique loops catch gaps. Every finding includes sample sizes and context.

Learns from every run

Run /deepwork deepwork_jobs learn after any job to capture what worked and improve the skill automatically.

Git-versioned everything

All work happens on branches. Skills are version-controlled. Roll back to prior versions, keep skills in sync across your team.

Composable skills

Skills can call other skills. Chain a competitive research skill into a comparison document skill — automate from raw research to final deliverable.

Example workflows

Define your own, or start from these. Each is a skill you build once and run whenever you need it.

Competitive Research

Track competitors weekly, generate diff reports, and produce strategic positioning recommendations.

Email Triage

Scan inbox, categorize, archive, and draft replies. Runs on a schedule.

Tutorial Writer

Turn expertise into documentation. Research the topic, draft, review against style guides, and polish.

Data Analysis

Connect to data, define research questions, query, validate findings, and produce a comprehensive report.

SaaS User Audit

Quarterly audit of who has access to which services. Find forgotten licenses and save money.

Founder Outreach

Research companies, understand their needs, and craft personalized outreach messages.

Get started

Claude Code (recommended)

# Install the plugin
/plugin marketplace add https://github.com/Unsupervisedcom/deepwork
/plugin install deepwork@deepwork-plugins
# Start a new session, then define your first job
/deepwork Make a job for doing competitive research

Other harnesses

DeepWork also supports Gemini CLI (partial support). OpenAI Codex and other harnesses are planned. See the GitHub README for the latest.

Try DeepWork

Free and open source. If you're already paying for a Claude Max subscription, each automation costs you nothing additional.

Questions? Open an issue or reach out at labs@unsupervised.com