AI Tip: Stop Writing One-Shot Summaries — Use Chain of Density Prompting

TL;DR

Chain of Density prompting produces AI summaries that rival professional human writers by iteratively packing more entity-rich information across 5 fixed-length passes. The sweet spot is iteration 3 — beyond that, readability drops. One prompt generates all 5 outputs, so you invest zero extra time.

If you've ever asked an AI to "summarize this article" and got back a generic, surface-level paragraph, you already know the frustration. The default summary is either too vague (missing key details) or too bloated (rambling on about tangents). The fix? Chain of Density (CoD) prompting — a research-backed technique from a 2023 paper by Adams et al. (arXiv:2309.04269) that systematically produces richer, denser summaries without increasing length.

What Is Chain of Density (CoD)?

Chain of Density is an iterative prompt engineering method. Instead of asking for one summary, you ask the AI to refine a summary across 5 fixed iterations, each time packing in more relevant information while keeping the exact same word count. The result: summaries that rival human-written ones in informativeness and readability.

The core insight is about entity density — the number of salient nouns and noun phrases per token. Human-written summaries naturally hit a sweet spot of ~0.15 entities per token. Most one-shot AI summaries either fall far below that (too vague) or exceed it chaotically (too dense). CoD walks the model methodically toward that sweet spot.

How It Works (The Prompt Template)

Here's the core CoD prompt template. Paste it into any capable LLM (GPT-4, Claude, Gemini) with your article:

Article: {{ YOUR_ARTICLE_TEXT }}

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.

Guidelines:
- The first summary should be long (4-5 sentences, ~80 words) yet highly non-specific.
- Make every word count: rewrite to improve flow and make space for additional entities.
- Use fusion and compression, removing uninformative phrases like "the article discusses".
- Never drop entities from the previous summary.

Answer in JSON: a list (length 5) of dicts with keys "Missing_Entities" and "Denser_Summary".

Why This Beats One-Shot Summarization

In the original paper, researchers compared CoD summaries to human-written ones and to GPT-4 using a vanilla "Write a VERY short summary" prompt. The results were striking:

• Step 3 of CoD matched the entity density of professional human-written summaries almost exactly.
• Human evaluators consistently preferred CoD outputs over one-shot summaries, rating them higher on informativeness without sacrificing readability.
• The iterative structure prevents the model from fixating on one angle — it's forced to balance breadth and depth across each pass.

The practical win: you spend zero extra time writing multiple prompts. Everything happens in a single call. The model does the iterative work internally (or across sequential calls if you prefer step-by-step control).

Real-World Use Cases

This technique isn't just for academic papers. Try it for:

• Research digestion: CoD-summarize 10 papers in parallel, then skim the Step-3 outputs for quick literature reviews.
• Meeting notes: Paste transcripts into the CoD template. The Step-3 output gives you an actionable, decision-dense summary in seconds.
• News briefing: Run articles through CoD each morning. The compressed, entity-rich summaries are faster to scan than headlines alone.
• Content repurposing: Take a long blog post, run CoD, and use the densest iteration as your social-media or newsletter excerpt.

Avoiding the Common Pitfall

The main trap with CoD is going too far. Steps 4 and 5 often decrease readability — the model starts shoehorning in minor entities that disrupt narrative flow. Stick to 2–3 iterations for the best balance of density and clarity. If you need more detail, run a second CoD pass focused on a specific subtopic rather than pushing past step 3.

What This Means for You

Here are three practical ways to apply CoD starting today:

• Use it as your default summarizer. Save the CoD template as a snippet in Claude/GPT/ChatGPT. Paste any article over 300 words and always use iteration 3 for the cleanest, densest output. This single change eliminates the "too vague" failure mode in one-shot summarization.
• Batch-research faster. Paste 10 papers, reports, or meeting transcripts into the CoD template in parallel. Skim the Step-3 outputs to triage which ones deserve a full read. At 5 iterations × 2 seconds per run, you get literature review-level summaries in ~10 seconds per paper.
• Repurpose content cross-platform. Run your latest blog post through CoD, then use iterations 1 (long form) for LinkedIn, 3 (dense) for your newsletter, and 5 (ultra-compressed entities) for Twitter threads. The same prompt produces three different formats from one source.

For the original paper and benchmarks, see Adams et al. — Chain of Density (arXiv) and Simon Willison's breakdown of the technique.

Getting Started Today

Copy the template above, drop in any article longer than 300 words, and watch the AI transform a vague blob into a gem of concentrated insight. For more quick productivity wins, check out our previous tip on AI debugging tools that actually work.

The bottom line: One-shot summarization is the "print debugging" of AI content work. Chain of Density is the structured, repeatable alternative that delivers research-grade results from a single prompt. Try it once, and you'll never go back.

← Back to all posts