If you’ve ever asked an AI to summarize a long article, you’ve probably experienced a common frustration: the summary is either too vague, misses the most important facts, or is padded with repetitive filler like, “This article discusses…”
It can feel like the AI just skimmed the text.
To fix this, prompt engineers use highly effective, overlapping concepts: the Rereading technique and the Chain of Density (CoD). By forcing the AI to iteratively reread the source material and progressively pack in more facts, you can transform a fluffy paragraph into a highly informative, gold-standard summary.
Here is a breakdown of how these concepts work and how you can use them.
The Core Problem: Single-Pass Reading
Large Language Models (LLMs) naturally predict the most likely next word. When you ask for a summary in a single prompt, the AI does a single “pass” over the text. This often results in:
- Lead Bias: The AI focuses heavily on the first few paragraphs and ignores the end of the document.
- Low Information Density: The AI uses too many words to convey too few actual facts.
Enter Chain of Density (CoD)
Introduced by researchers in the paper “From Sparse to Dense: GPT-4 Summarization with Chain of Density Prompting” in 2023, Chain of Density is a prompt engineering technique designed specifically to fix bad summaries. It forces the AI to write a summary, evaluate it, and rewrite it multiple times—without increasing the word count.
Where “Rereading” Fits In
The Chain of Density relies entirely on a structured rereading loop. (Note: In AI prompting, “rereading” can also refer to the RE2 technique, where you explicitly prompt the AI to “Read the text again” to improve logical reasoning. CoD applies this same philosophy to summarization). During the CoD process, the AI must:
- Read the source text and write a basic summary.
- Reread the source text to identify 1 to 3 specific entities (names, numbers, locations, concepts) that it missed.
- Rewrite the summary to include the new entities while cutting out filler words to maintain the exact same length.
- Repeat this loop 5 times.
A Concrete Example
Let’s look at how a summary evolves through this rereading and densification process.
The Source Text is a 500-word news report about a new solar-powered train launching in Byron Bay, Australia, costing $4 million, reaching speeds of 50 km/h, carrying 100 passengers, and built by the Byron Bay Railroad Company.
- Iteration 1 (Sparse & Fluffy):“This article discusses a new solar-powered train that has recently launched in Australia. The train is designed to be environmentally friendly and runs on renewable energy. It will carry passengers along a coastal route, offering a sustainable transport option for locals and tourists.” (Word count: 43)
Critique: Standard AI fluff. It misses the specific location, the cost, the speed, and the company. - Iteration 3 (The “Goldilocks” Zone):“The Byron Bay Railroad Company has launched a new solar-powered train in Australia. Running entirely on renewable energy, the $4 million train carries up to 100 passengers along a coastal route. It provides a sustainable transport option for locals and tourists.” (Word count: 41)
Critique: Much better. By rereading the text, the AI found missing entities (Byron Bay Railroad Company, $4 million, 100 passengers) and compressed its own language to make room for them. - Iteration 5 (Highly Dense):“Byron Bay Railroad Company’s $4 million solar-powered train, carrying 100 passengers at 50 km/h, launched in Australia. This sustainable, renewable-energy coastal transport serves locals and tourists, eliminating fossil-fuel reliance.” (Word count: 30)
Critique: Extremely packed. Research shows that humans usually prefer Iteration 3 or 4, as Iteration 5 can sometimes feel a bit too dense or robotic to read comfortably.
How to Try It Yourself
You don’t need to be a developer to use this. You can trigger this exact behavior by copying and pasting the following framework into your next prompt:
Article: [Insert your text here]
You will generate increasingly concise, entity-dense summaries of the above article. Repeat the following 2 steps 5 times:
Step 1: Reread the article and identify 1-3 informative entities (specific people, numbers, places, concepts) that are missing from your previously generated summary.
Step 2: Write a new, denser summary of identical length (about 50 words) which covers every entity from the previous summary PLUS the new missing entities.
Guidelines:
- Make every word count. Compress and rewrite sentences to make space for the new entities.
- Remove uninformative phrases like “The article discusses”.
- Output the result as a numbered list of 5 summaries, from Iteration 1 to Iteration 5, so I can choose the best one.
That is a fantastic, highly practical workflow. In fact, adding a dedicated “find themes” step is a brilliant way to upgrade the standard Chain of Density technique.
If an AI just packs in random facts to make a summary “dense,” the result can feel disjointed. By forcing the AI to extract the themes first, you ensure that the final dense summary is actually anchored to the true narrative of the text.
The “Theme-Driven Density” Prompt
You can copy and paste this exact template whenever you need to process a document.
Text to Analyze: > [Paste your text here]
Task: Follow this 2-step workflow to process the text above.
Step 1: Find Themes (The Rereading Phase)
Read the text carefully. Identify and list the 3 core themes or main arguments. For each theme, explicitly state 1 or 2 concrete entities (specific data points, names, dates, or key concepts) from the text that support it.
Step 2: Write Summary (The Density Phase)
Using only the themes and entities you extracted in Step 1, write a highly informative, single-paragraph summary (roughly 50 to 75 words).
Constraint 1: Do not use filler phrases like "The author notes" or "This text discusses."
Constraint 2: Ensure every specific entity from Step 1 is woven into the text. Make every word count.Why This Workflow Succeeds
By structuring your request this way, you are manipulating how the AI processes information under the hood:
- It prevents hallucinations: By forcing the AI to list the themes and supporting facts before it writes the summary (Step 1), you ground its reasoning. It is much less likely to make things up when it has an outline to follow.
- It naturally creates density: In Step 2, the AI is constrained. It has a high volume of specific information (the themes and entities) and a strict word limit. This forces it to naturally adopt the Chain of Density writing style, combining sentences and dropping fluffy adjectives.
- It gives you an audit trail: If the final summary feels slightly off, you can look at the “Themes” list from Step 1 to see exactly where the AI misunderstood the source text.