What is chain of density prompting?
Chain of density prompting is a summarization technique where a model rewrites a summary several times in a row, each iteration more information-dense than the last, without getting longer. You start with a loose, readable summary and end with a compact one packed with specifics. The length stays roughly fixed; the information per word increases.
The term comes from a 2023 paper by Adams et al. published on arXiv, which studied how GPT-4 could produce summaries that were both shorter and more informative by iterating over the same source text. The researchers found that mid-density summaries were often preferred by human readers over both the sparsest and most compressed versions, a useful calibration point when you are deciding how many iterations to run.
For a team lead, the practical value is straightforward. Long reports, meeting transcripts, research briefs, and customer feedback threads all need to be condensed before they can be acted on. Chain of density prompting gives you a repeatable way to do that across ChatGPT, Claude, Copilot, or Gemini, without rewriting your prompt from scratch every time.
It sits within the broader family of prompting techniques that shape model output by structuring the task rather than by changing the model itself.
How it works
The core mechanic is iteration. You give the model a passage to summarize, then ask it to produce a series of increasingly dense rewrites, each one adding a fixed number of new entities (concepts, names, claims) while keeping the word count roughly constant.
The original technique, described by Adams et al. (2023) in "From Sparse to Dense: GPT-4 Summarization with Chain of Density Prompting", used five iterations and a target of two to three new entities per step. The summary gets shorter in feel even though it stays the same length, because filler phrases get compressed to make room.
Each rewrite has two constraints:
- Length holds steady. The model cannot add words to accommodate new content. It has to cut something to fit something in.
- Nothing already present can be dropped. Every entity from the previous version must survive. Information only accumulates.
Those two rules together are what drives densification. The model is forced to find tighter syntax, merge related clauses, and eliminate anything that does not carry a fact.
In practice, you run this as a single prompt with the full chain of instructions written out, rather than sending five separate messages. That keeps the model's context consistent across all iterations and reduces drift between steps.
Worked examples
Each example below shows the same prompt run through two or three density passes. The actual model outputs are marked [NEEDS REAL OUTPUT] and must be filled before publication.
Example 1: Summarizing a research report
Pass 1 prompt
Summarize this quarterly market report in 3-4 sentences.
[PASTE REPORT TEXT]Pass 1 output: [NEEDS REAL OUTPUT]
Pass 2 prompt
Rewrite that summary in the same length. Add two specific findings
you omitted, without removing anything already there.Pass 2 output: [NEEDS REAL OUTPUT]
Commentary: The second pass forces the model to compress rather than expand. Readers get more signal in the same space.
Example 2: Condensing a product update email
Pass 1 prompt
Summarize these product release notes in 4 sentences for a non-technical audience.
[PASTE NOTES]Pass 1 output: [NEEDS REAL OUTPUT]
Pass 2 prompt
Revise that summary to the same length. Include the two most important
feature names, without dropping any existing detail.Pass 2 output: [NEEDS REAL OUTPUT]
Commentary: Named features replace vague phrases like "new capabilities," making the summary actionable.
Example 3: Tightening a meeting recap
Pass 1 prompt
Summarize this meeting transcript in 5 sentences.
[PASTE TRANSCRIPT]Pass 1 output: [NEEDS REAL OUTPUT]
Pass 2 prompt
Rewrite that recap at the same length. Add the owner and deadline
for each action item without removing current content.Pass 2 output: [NEEDS REAL OUTPUT]
Commentary: Accountability details that belong in every recap are often dropped in a single-pass summary. The density pass retrieves them.
When to use it and when not to
Chain of density prompting earns its place in a specific situation: you have a long source document and you need a summary that a busy reader can actually act on, without losing the details that matter.
It works well for:
- Research digests and literature reviews. Each pass pulls in another finding without expanding the word count.
- Executive summaries of reports. You can target a precise length and hit it reliably.
- Meeting notes and call transcripts. Cutting filler while keeping decisions and owners intact is exactly what the technique is designed for.
- Legal or compliance documents where removing a clause accidentally changes meaning. The iterative structure lets a reviewer catch drift early.
Where it does not help, and this is worth being direct about: if your source material is thin, the process produces diminishing returns fast. Running five compression passes on a 200-word email wastes time and sometimes creates summaries that feel strangled rather than tight.
It also adds friction when speed matters more than precision. A single zero-shot prompt asking for a "three-sentence summary" will often be good enough, and it takes thirty seconds instead of five minutes.
For quick one-off tasks, simpler approaches are the right call.
Chain of density prompting compared with related techniques
Chain of density prompting is one of several iterative techniques, and knowing which to reach for matters.
| Technique | Core mechanic | Best for | Weakness |
|---|---|---|---|
| Chain of density | Repeated compression, adding entities per pass | Summarization, tight briefs | Requires multiple passes; slower |
| Chain of thought | Step-by-step reasoning visible in output | Math, logic, multi-step decisions | Produces longer outputs, not shorter |
| Zero-shot prompting | Single instruction, no examples | Quick tasks, first drafts | Less control over density or structure |
| Iterative refinement | Free-form rewriting based on feedback | Open-ended editing | No systematic constraint driving improvement |
Chain of thought asks the model to show its working. Chain of density asks it to hide the working and compress the result. They solve opposite problems, which is why combining them is rarely useful: you get a dense output that no one can audit.
Zero-shot prompting is faster but gives you no systematic way to control information density. If the first summary is too thin or too padded, you are left guessing at the right instruction to fix it. Chain of density sidesteps that guessing by making compression an explicit, repeatable step.
Iterative refinement is the closest relative. The difference is constraint: chain of density gives the model a specific rule (same length, more entities) rather than a vague instruction to "improve" the text. That rule is what makes the outputs consistent enough to reuse as prompt templates.
Common mistakes
The most frequent error is treating the first compressed version as good enough. Chain of density prompting produces value through iteration. Running one compression pass and stopping means you have rewritten a summary, not refined one. The compression-then-review cycle is the actual technique.
A related problem: compressing without removing anything. Writers instinctively reach for shorter phrases and tighter verbs, which helps, but true density requires cutting entities, not just words. If your third iteration is still mentioning everything your first one mentioned, the density hasn't increased. Something has to go.
Watch the readability floor. Some teams push too many iterations and end up with a summary that is accurate but genuinely hard to parse. A sentence that requires two readings to understand is not a better sentence. When reviewers start flagging output as "dense" in a negative sense, that's a signal to stop one iteration earlier.
One practical pitfall specific to teams: inconsistent stopping points. One person runs two iterations, another runs five, and the outputs are hard to compare or combine. Agreeing on a target density level before you start, rather than stopping when it "feels right," produces more consistent results across different writers and different models.
Finally, avoid applying this technique to text that is already short. A 50-word paragraph does not benefit from chain of density prompting. The technique earns its cost on longer source material.
Using this across a team
Solo use of chain of density prompting is straightforward. The harder problem is consistency: when five people on the same team all run their own versions of a CoD prompt, you get five different summary styles, five different definitions of "entity," and five different ideas about what "concise" means.
The fix is a shared prompt template. Agree on the number of iterations (three is a reasonable default for most summaries), lock in the entity definition, and save that as a team prompt rather than leaving everyone to improvise from memory. When the template lives in a shared library, anyone can run it without reinventing the setup, and edits to the template propagate to the whole team at once.
That shared baseline also makes quality review easier. If two people summarize the same source document and get meaningfully different results, you can trace the divergence to a prompt difference rather than guessing. Reviewing outputs becomes a conversation about the prompt, not about individual judgment.
One practical note: document which iteration you're handing off. If someone compresses a second-iteration summary to save time, the reader may not realize density has already been applied once. A short label, "CoD pass 1 of 3" or similar, costs nothing and prevents confusion downstream.
Frequently asked questions
What is chain of density prompting?
Chain of density prompting is a technique where you ask a model to produce multiple versions of a summary, each one shorter than the last but containing the same essential information. The idea is to compress meaning progressively rather than asking for a single summary and accepting whatever the model decides to cut.
How is chain of density prompting different from just asking for a short summary?
A single summary request gives you one output shaped entirely by the model's judgment about what matters. Chain of density prompting forces the model to make that compression visible across several steps, which lets you intervene, compare versions, and catch information that was dropped prematurely.
Which AI models support chain of density prompting?
Any model that can follow multi-step instructions works, including ChatGPT, Claude, Copilot, and Gemini. The technique relies on instruction-following ability, not on any model-specific feature. Results vary across models, so it is worth running your prompt on whichever tool your team uses most.
Does chain of density prompting work for languages other than English?
It works in most languages the underlying model supports, though compression quality tends to be less reliable in lower-resource languages. If your team works multilingually, test on a representative sample before rolling it out at scale.
When should I not use chain of density prompting?
Skip it when you need a first draft rather than a condensed version of existing material. Chain of density prompting is a summarization tool, not a generation tool. For original content, zero-shot prompting or a structured system prompt will serve you better.
Further reading
The original research behind chain of density prompting was published in the 2023 paper "From Sparse to Dense: GPT-4 Summarization with Chain of Density Prompting" by Adams et al. It is a short read and the appendix includes the exact prompts the researchers used, which are worth comparing against your own iterations.
For broader context on how chain of density fits alongside other iterative approaches, the Convergence guide to prompting techniques covers the major families in one place. If you are new to structuring instructions more generally, the guide to system prompts is a practical starting point before you add density iterations on top.
Two adjacent techniques come up often in the same conversations. Zero-shot prompting is the baseline worth understanding before you add iterative refinement, because it shows you what a model can do with a single, well-formed instruction. Once you see where single-pass prompts fall short on compression tasks, the case for iteration becomes obvious.
If you want pre-built prompts your team can start from without writing the iteration chain from scratch, the Convergence prompt template library includes summarization and condensation templates formatted for ChatGPT, Claude, Copilot, and Gemini.
Finally, Ethan Mollick's Substack, One Useful Thing, covers applied AI use in a similarly plain-spoken register and is worth following for broader context on how models behave in professional settings.