I use AI tools to summarize things all the time: the transcript of a panel I missed but recorded, a long report that I wanted to see if it was worth reading, show notes for a podcast I can barely remember after all the editing—it's really one of the most useful things that AI tools can do. 

But, I'm sure you've run into this too: often, AI tools just don't quite capture everything relevant on the first go. Turns out that a team of researchers from Columbia, Salesforce, and MIT have been struggling with summarizing too, and they created a new technique to help: it’s called "Chain of Density." Here’s how it works and how to use it for your own summaries.

TL;DR: Simplified Chain of Density prompt for better summaries

Prompt 1: Vanilla Summary

Please write a summary of the attached transcript. Make sure you include all the important points in the article. The summary should be about 200 words.

Prompt 2: Missed Points

Please review the transcript and identify all important points that were missed in the summary.

Prompt 3: Importance

Can you take your list of points and rate how important they are to the overall transcript?

Prompt 4: Rewrite

Please incorporate [those points/the points you rated as Very Important] into the summary. Make sure that all the points in the original summary are also included. The resulting summary should be 200 words in length.

Why is density important for summaries?

The premise behind Chain of Density is that summaries are tough to get right. By definition, a summary should be denser than the original if it's going to keep all relevant details in a much shorter format. But that's the problem: making it denser can make it hard to read. It needs to be the Goldilocks of information—not too sparse and not too dense.

That means details need to be handled in a systematic way to keep the summary at the right density. Summarization can be managed by abstracting, compressing, and fusing concepts instead of mentioning each specific one. But try to compress it too much, and it will lose the key components that make it up.

You can take out some details without changing the overall meaning much, but taking out other details can result in a big change. That is what makes summarization so fraught—how do you make sure the summary is faithful to the original? 

Luckily, Chain of Density has a potential answer for you.

What is Chain of Density?

Remember when you had a 10,000 word term paper due and you tried to add filler words to hit the word count without actually having to research and write new content? Just me? I majored in math, after all.  

Anyway, that's what adding sparseness does: it adds length without actually adding any content. In AI research, density is measured by "entities," which is kind of like a shortcut for "detail." For instance, "Math major Briana Brownell struggled writing college term papers" is much denser than " Briana Brownell, who is a math major, consistently and repeatedly experienced significant and considerable difficulty and hardship whenever she found herself faced with the demanding and arduous task of writing her college term papers." 

So here's the Chain-of-Density process. First, you have the AI tool create a deliberately sparse summary. Then, you ask the tool to iteratively include more and more "identified entities" (read: details) without making the summary longer. This makes the summary denser.

At some point, you reach a level where readability starts to suffer because the summary is too dense. Pass that point, and you eventually max out in density where you can't add more detail without increasing the length or dropping details you had previously added.

You really need to figure out how to hit that sweet spot. In the research paper, they tried the technique with news articles from CNN and the Daily Mail. The team ran these articles through five steps of "densification," each of which attempted  to add 1–3 new identified entities. (I've included their prompt template and an example at the end of the article.) 

Initially, the summaries were extremely sparse and uninformative, but eventually, they got denser and captured all of the article's main points. The sweet spot was about three steps of densification, which captured about 5–9 details in the 70 word summary,, so about one detail per 10 words or so.  This made the summary about as dense as a human-created summary, but denser than a vanilla ChatGPT prompt.

The original Chain of Density prompt

Here’s the prompt the researchers used in their original paper. While it worked for 500-word news articles, it didn’t quite work for the longer pieces I usually like to summarize, and I’ll explain how I tweaked it in the next section.

Article: [link] (or attach a PDF)

You will generate increasingly concise, entity-dense summaries of the above Article. 

Repeat the following 2 steps 5 times. 

Step 1. Identify 1-3 informative Entities (";" delimited) from the Article which are missing from the previously generated summary. 

Step 2. Write a new, denser summary of identical length which covers every entity and detail from the previous summary plus the Missing Entities. 

A Missing Entity is: 
- Relevant: to the main story. 
- Specific: descriptive yet concise (5 words or fewer).
- Novel: not in the previous summary. 
- Faithful: present in the Article. 
- Anywhere: located anywhere in the Article. 

Guidelines: 
- The first summary should be long (4-5 sentences, ~80 words) yet highly non-specific, containing little information beyond the entities marked as missing. 
- Use overly verbose language and fillers (e.g., "this article discusses") to reach ~80 words.  
- Make every word count: rewrite the previous summary to improve flow and make space for additional entities. 
- Make space with fusion, compression, and removal of uninformative phrases like "the article discusses". 
- The summaries should become highly dense and concise yet self-contained, e.g., easily understood without the Article. 
- Missing entities can appear anywhere in the new summary. 
- Never drop entities from the previous summary. If space cannot be made, add fewer new entities. 

Remember, use the exact same number of words for each summary. 

My simplified Chain of Density prompt

Like me, you might be summarizing different kinds of things, and Chain of Density is fairly flexible if you make a few changes to the technique. You can see my prompt process down below.

I tried it for the following things:

• A webinar transcript (~10,000 words)
• A 11-page report (~5,000 words)
• An article (~2,000 words)

The original prompt didn’t work for me for a few reasons. 

For one, these pieces are much longer than a 500-word news article, and I found that the super-short 70-80 word summary just wasn't long enough to capture anything other than the surface information.

Second, trying to get all five summaries in one prompt took me past the context window when I was doing longer summaries. There's an easy workaround though: I just changed the prompt style up to be an interactive step-by-step—that is, I entered each prompt one at a time instead of all at once. It took longer but ended up giving me better results.

Because of this, I kept the Chain of Density method philosophy but streamlined the prompt to make it more succinct. I started by asking for a general summary, then added density from there. I asked the tool to identify important points that were missed in the summary, and then prompted it to include them in a rewritten summary. I found this worked much better than trying to get it to generate a super sparse summary and adding density slowly, which, for longer summaries, was difficult to do (turns out there's a maximum to how much you can pad text—something I'm sure any teacher marking essays already knew).

Here’s my simplified prompt:

Prompt 1: Vanilla Summary
Please write a summary of the attached transcript. Make sure you include all the important points in the article. The summary should be about 200 words.

Prompt 2: Missed Points
Please review the transcript and identify all important points that were missed in the summary.

Prompt 3: Importance
Can you take your list of points and rate how important they are to the overall transcript?

Prompt 4: Rewrite
Please incorporate [those points/the points you rated as Very Important] into the summary. Make sure that all the points in the original summary are also included. The resulting summary should be 200 words in length.

My tests showed that the Vanilla prompt was usually pretty good already, but after I got the prompt structure working, I was impressed. Even though this technique was more work than just asking for a summary, it definitely improved the quality of the summary when more ideas were added.  

As the researchers identified, hitting that sweet spot between too sparse and too dense was tricky. When I tried to add too many additional points, it ended up just reading like a laundry list of buzzwords and the readability tanked. My suggestion is to add just a few new points at a time. If you want, you can just choose which points you think are most important to add to the summary, or ask ChatGPT to help you prioritize which ones to add.

If you're doing a lot of summarization, it's definitely worth it to try mastering this technique. It can turn a good summary into a great one — if you can hit that perfect balance between detail and readability.

Appendix: A few of my results

Vanilla Summary

The webinar "AI for Today's Boardroom" by BoardProspects, hosted by Mark Rogers, features Lake Dai, an AI expert and venture investor. The discussion focuses on the impact of AI on various industries and the role of board members in navigating this transformation.

Dai explains that AI, though not a new concept, has significantly advanced due to increased data, computational power, and innovative algorithms. She differentiates between AI, machine learning, and deep learning, and clarifies generative AI and AGI (Artificial General Intelligence).

Boards need to understand AI's potential to disrupt industries and drive growth. Dai emphasizes the importance of board members asking the right questions about AI's impact, data integrity, safety, and legal implications. She discusses the practicality of having AI experts on boards or integrating AI oversight into existing committees, like audit or risk committees.

Use cases of AI in traditional industries, such as restaurants and jewelry design, illustrate AI's broad applicability. The conversation also touches on regulatory challenges, the need for AI ethics, and the risks of AI, including biases and data privacy concerns.

Dai highlights the future of AI, including the potential for specialized GPT models for different sectors and the necessity for distributed computing to manage energy consumption and efficiency. She concludes by discussing the evolving landscape of AI governance and the critical role of AI ethics and alignment with human values.

This session underscores the transformative potential of AI and the critical need for board members to stay informed and proactive in addressing AI-related opportunities and risks.

Dense Summary

The webinar "AI for Today's Boardroom" by BoardProspects, hosted by Mark Rogers, features AI expert Lake Dai. It addresses AI's impact on industries and boardroom governance. Dai explains AI's evolution, differentiating between AI, machine learning, and deep learning, and clarifying concepts like generative AI and AGI.

Boards should ask key questions about AI's disruption potential, strategic alignment, investment, data integrity, safety, and legal implications. Use cases in traditional industries, such as restaurants using AI for marketing and jewelry design, demonstrate AI's broad applicability.

Dai emphasizes AI ethics, safety, and alignment with human values, drawing an analogy to co-parenting. Challenges with pre-trained models like ChatGPT include cost, flexibility, and hallucination, prompting shifts to open-source and managed open-source solutions.

She discusses the potential for vertical AI models (e.g., healthcare GPT) alongside large foundational models. The webinar highlights the energy consumption conflict of AI data centers and corporate sustainability goals, suggesting a boost to clean energy innovation.

Federal and state AI regulations are evolving, with California leading, impacting sectors like finance and healthcare. Boards might need AI experts or committees. The insurance industry is expected to develop products for AI risks.

Dai also notes the potential for AI to enhance government customer service, balancing efficiency with data privacy and security concerns. The session underscores AI's transformative potential and the need for proactive board engagement.