Deepwork — Claude Code as your Personal Assistant

“I used to spend an hour going back and forth to get a few pieces of content. With Deepwork I do 5x that and waste much less time on revisions.”

“I create a job, let it run in the background, and feel assured I’ll get great quality output when I check back.”

“Claude Code alone requires a LOT of steering to ensure quality. Deepwork gives it the power to do repetitive tasks at scale with rigorous checks.”

Same User, Same Prompt

Zero-shot data analysis

Without Deepwork

With Deepwork

Deepwork added 5 extra pages, $8.4M revenue impact, ROI estimates, and data quality flags

The agent worked 3x longer on its own — 30 min vs 10 min

Both runs used the same prompt: “Go look at the telco data in the database (use the DSN env var to connect) and build a PDF report about what is driving Customer Churn. Put it in the demo_analytics_reports folder.”

No further iteration — this is the direct, unedited result. Claude connected to the database, ran queries, and produced a 5-page report with an executive summary, bar charts of churn drivers, customer segment breakdowns, and a page of strategic recommendations.

Agent ran autonomously for ~10 min

Run through a Deepwork data analysis job that took ~10 minutes to create. The agent spent 3x longer exploring the data, self-critiquing its findings, and iterating on presentation — producing a 10-page report with detailed charts, financial impact estimates, and actionable recommendations.

Agent ran autonomously for ~30 min

	Without	With Deepwork
Key findings	6	7
Churn drivers analyzed	10	8
Revenue impact quantified	✗	$8.4M
Compound risk segments	✗	Identified
ROI per recommendation	✗	Estimated
Data quality flags	✗	4 issues noted

First time making a Manim animation

Without Deepwork

With Deepwork

Deepwork added no manual editing, structured narration, scene planning & QA, and an end-to-end pipeline

The agent worked 5x longer on its own — 240 min vs 45 min

Using Adithya Kolavi's Manim Skill

The user loaded a Manim skill and provided links about their company along with a custom prompt requesting educational, viral, and visually-engaging ideas. Claude used the skill to generate the animation, and then the user paired with Claude for 60–90 minutes of manual iteration, suggesting edits to the video as they went. The result was a solid proof of concept.

Agent ran autonomously for ~45 min

Using the Deepwork Manim job

A Deepwork job was defined in approximately 10 minutes to produce Manim animations in the style of 3Blue1Brown. The output was then run through 5 iterations of the learn command, which diagnosed issues and updated the job automatically — no video was ever manually edited. The only input was a URL and a link to product documentation.

Agent ran autonomously for ~240 min

Process beats reference material

We asked a separate Claude Code instance to compare the Deepwork job for Manim Animation vs Adithya Kolavi’s Manim skill. Despite having no hard-coded Manim library reference, a well-defined process produced better results.

For novice users, skills that have stronger process definitions lead to better results than skills that provide more reference material as context.

Deepwork job vs. Adithya Kolavi's Manim skills — scored by an independent Claude Code instance

Category	manim-composer	Manim-CE	Manim-GL	Deepwork
API & library reference (4)	1	3	3	0
Advanced rendering (4)	0	2	3	0
Planning & composition (3)	2	0	0	3
Production & delivery (6)	0	0	0	6

Capability	manim-composer	Manim-CE	Manim-GL	Deepwork
API & library reference
ManimCE API reference	✗	✓ (23 rules)	✗	Partial
ManimGL support	✗	✗	✓ (18 rules)	✗
Working examples	✗	✓ (7)	✓ (75)	✗
Templates	✓ (1)	✓ (3)	✓ (4)	✗
Advanced rendering
3D guidance	✗	✓	✓	✗
Graphing/axes	✗	✓	✗	✗
Updaters/ValueTracker	✗	✗	✓	✗
Interactive dev (-se)	✗	✗	✓	✗
Planning & composition
Scene planning	✓	✗	✗	✓
Narration/TTS	✓	✗	✗	✓
Visual QA	✗	✗	✗	✓
Production & delivery
End-to-end pipeline	✗	✗	✗	✓
Peer review	✗	✗	✗	✓
Brand integration	✗	✗	✗	✓
Production pitfalls	✗	Partial	Partial	✓
ffmpeg assembly	✗	✗	✗	✓
Timing alignment	✗	✗	✗	✓

Bottom line: The three community skills are API reference libraries — what to use and how to structure scenes. The Deepwork job is an execution pipeline — taking a concept from idea to finished video, with narration, QA, and assembly built in.

Primary Research, Feb 2025 | www.deepwork.md

Example jobs

Each job is a multi-step workflow where Claude validates its own output at critical moments — so you get a finished result, not a first draft.

Blog Post from a Topic

Give it a topic and get back a publish-ready post. Claude researches, outlines, writes, reviews it from your reader’s perspective, and optimizes for SEO.

7 steps, checks itself twice

Product Research from a URL

Point it at a product and get a thorough research report. Claude browses the web autonomously, makes 10+ passes, and won’t stop until every section has real substance.

3 steps, checks itself 3 times

Data Analysis from Your Warehouse

Connect a data source and get a presentation-ready analysis. Claude explores the data, takes notes, then critiques its own findings from six different angles before delivering.

6 steps, checks itself twice

Email Campaign from Your Product

Describe your product and audience, get a full 14-day email sequence. Claude writes each email, then stress-tests every one for tone, persuasion, and clarity.

6 steps, checks itself 5 times

Competitive Intel from Live Websites

Run it monthly and get a report on what your competitors changed. Claude visits their sites, captures screenshots as evidence, and flags what’s different since last time.

4 steps, checks itself 4 times

Slide Deck from a Brief

Give it a topic and get a full presentation. Claude structures the argument, builds every slide, then reviews the whole deck for clarity and flow before you see it.

10 steps, checks itself 3 times

Daily Email Triage from Gmail

Run it on your inbox every morning. Claude labels, archives, and drafts replies in your voice — handles 100+ emails without losing context.

2 steps, checks itself twice

Image Generation with Self-Critique

Describe what you want and Claude generates it, looks at its own output, and refines it automatically before asking if you want further changes.

2 steps, checks itself once

Jobs can chain together — one job’s output becomes the next job’s input.

Tips to get the most out of Deepwork

01

Answer the questions thoughtfully

Run /deepwork_jobs and select define. DeepWork will ask 3–4 rounds of scoping questions.

Spending 10 minutes here saves hours later.

02

Give DeepWork relevant context before you create the job.

DeepWork reviews your thread context before asking questions. Pre-load the conversation with relevant context — links, folders, topics to do web searches on, examples of related work, diagrams, images — and the questions will be sharper and the output better.

We like to open with “we’re going to discuss automating a task today, can you read the following and let me know when you’re ready?” then paste everything in.

03

Set aside 30 minutes for your first run, run through a full define, execute, and learn loop.

Run /deepwork_jobs and select define. Answer the questions and Claude will generate a job.yml with a full workflow and a skill file for each step. This takes ~10 min. Don’t worry about reading the intermediate output (job.yml, step files, etc.) — it’ll make more sense after you’ve seen the full workflow in action.

Then run the main skill (e.g. /deepwork_jobs, not /deepwork_jobs.define). Let the agent complete the entire flow.

When it finishes, share your feedback and ask Claude to run deepwork learn skill to incorporate it. This full loop — define, execute, learn — is the best way to understand what DeepWork does.

Turn Claude Code Skills into Claude Code Workflows