Communicating Insights Visually

Anthropic published an article on How AI is transforming work at Anthropic.

We wanted to see whether LLMs can suggest visualization ideas to illustrate the ideas in the paper.

To generate this, we converted the article to Markdown and used the following prompt:

This article is about how AI is transforming work at Anthropic. Suggest how the following engineer productivity patterns can be illustrated using interactive animated charts, graphs, or infographics.

Specifically show the before-vs-after impact of AI on these engineer productivity patterns:

  1. Engineers are getting a lot more done, becoming more "full-stack" (able to succeed at tasks beyond their normal expertise)
  2. Engineers are accelerating their learning and iteration speed (tighter feedback loops)
  3. Engineers are tackling previously-neglected tasks.

For each pattern, I want 3 different animated graphs.

Render this as a single HTML file with tabs for each productivity pattern, and a gallery of animated charts for each pattern.

Here is the article:

{insert article here}

... with the top AI chatbots / coding agents. Here are the results and chats, sorted by quality:

Rank Quality AI Tool Chat
1 🟢 Gemini Canvas 💬 Chat
2 🟢 Gemini Canvas v2 💬 Chat
3 🟡 Claude Artifacts 💬 Chat
4 🟡 ChatGPT Canvas 💬 Chat
5 🟡 Claude Code 💬 Chat
6 🟡 Codex CLI 💬 Chat
7 🔴 Grok 4.1 Beta 💬 Chat
8 🔴 GitHub Copilot + Grok Code Fast 1 💬 Chat

Then, we extended it with:

Look closely for errors in your output and fix them.

Extend to 3 more animated graphs each. Be more diverse. Your aim is to give me creative ideas. Xenographics are welcome. Novel animation / interaction styles, artistry, xenographics, and diverse chart types are encouraged.

For the local coding agents, we added this line:

Use CDP at localhost:9222 to load the page and check for console errors, screenshots, etc. The page can be loaded from http://localhost:8000/...

Rank Quality AI Tool Chat
1 🟢 Gemini Canvas 💬 Chat
2 🟢 Gemini Canvas v2 💬 Chat
3 🟢 Claude Artifacts 💬 Chat
4 🟡 ChatGPT Canvas 💬 Chat
5 🟡 Claude Code 💬 Chat
6 🟡 Codex CLI 💬 Chat
7 🔴 Grok 4.1 Beta 💬 Chat
8 🔴 GitHub Copilot + Grok Code Fast 1 💬 Chat

The chatbots (ChatGPT, Gemini, Claude, Grok) took 5-10 minutes to generate the code. AI coding agents (Codex CLI, Claude Code, OpenCode) took as long has half an hour.

Cost-wise, Claude Code cost $1.91 and the rest might be comparable or lower.


Some of the ideas, especially by Gemini 3 Pro and Claude, are quote creative!

Throughput Flow: Visualizing commits as particles. From a trickle to a steady stream
Knowledge Neural Net. Visualizing an engineer's mental model. Isolated clusters connect into a dense mesh.
Repo Coverage Heatmap. Grid representing codebase files. Dark spots (untouched legacy code) light up.
Dark Matter Radar. Detecting invisible/ignored tasks
Feedback Pulse. Frequency of iteration heartbeats
Iteration Spiral. More cycles completed in same time period
Expertise Growth Rings. Cumulative skill layers added with AI assistance
Prompt→Ship Token Flow (pipeline kinetics). Tokens represent iterations. After = faster tokens + fewer backtracks. Pipeline kinetics: tokens represent iterations; after = faster + fewer backtracks
Workbench Blocks (polyomino task stacking). Metaphor: the same day fits more “weird-shaped” tasks when AI trims edge friction. Xenographic: polyomino tasks stack into domain bins (more variety fits after AI)
Quality-of-Life Impact Grid. Colored cards based on lower - higher impact

My expertise as a data visualization expert has two parts:

  1. Suggesting innovative visualisations
  2. Picking the right one for the job

AI does a better job of #1 than I do right now.