# LLMs Relearn Telegraphese in 1866 Style

> Published 2026-10-07 · https://www.promptzone.com/deepa_kowalski/llms-relearn-telegraphese-in-1866-style-ehk

A [Hacker News thread](https://fiveminutesforward.com/post/2026-10-04-telegraph-test/) posted on October 4, 2026, reports that LLMs produce shorter, cheaper outputs when prompted to mimic 1866 telegraph style.

The experiment tests whether models can drop modern prose habits and adopt the abbreviated, high-signal format used in paid-by-the-word telegrams.

## What Telegraphese Prompts Do

Telegraphese strips articles, conjunctions, and polite phrasing. Sentences become subject-verb-object sequences with numbers and abbreviations.

Prompts instruct models to output only essential facts, omit transitions, and use 19th-century abbreviations such as "rec'd" or "fwd".

Early tests show token counts drop 35-48% on summarization tasks while preserving key entities and numbers.

## Token and Cost Numbers

Standard GPT-4o summaries of 500-word articles average 620 tokens. The same articles rewritten in telegraphese average 340 tokens.

Claude 3.5 Sonnet shows a 41% reduction. Llama 3.1 70B records a 37% cut.

At current API rates this translates to roughly $0.0009 saved per summary on GPT-4o.

## How to Try It

Use this base prompt template:

"Rewrite the following text in 1866 telegraphese. Omit all articles, conjunctions, and filler. Keep only names, dates, numbers, and actions. Output one line per fact."

Append the source text and set temperature to 0.2 for consistency.

Community nodes for ComfyUI and LangChain already include a "telegraph" formatter node released last week.

## Pros and Cons

- Pros: Lower inference cost, faster API responses, easier downstream parsing.
- Cons: Reduced readability for non-technical users, loss of nuance in legal or medical text, occasional abbreviation collisions.

## Alternatives and Comparisons

| Style          | Token Reduction | Readability Score | Best For                  |
|----------------|-----------------|-------------------|---------------------------|
| Telegraphese   | 35-48%         | 62/100            | Logs, alerts, summaries   |
| Bullet points  | 22-30%         | 78/100            | Reports, notes            |
| Standard prose | 0%             | 91/100            | Customer-facing content   |

Compared with chain-of-density or tree-of-thoughts, telegraphese requires no extra reasoning steps and works on any base model.

## Who Should Use This

Developers building monitoring dashboards or cost-sensitive chat agents benefit most. Skip it for customer support bots or any output read by non-experts.

Researchers studying token efficiency can treat it as a zero-training compression technique.

> **Bottom line:** Telegraphese prompting delivers measurable token savings on current models without fine-tuning or new infrastructure.

Early HN comments note the method revives an old compression trick rather than inventing a new one. The discussion reached 70 points and 46 comments within 48 hours.

The approach fits existing prompt pipelines immediately and requires only a single system prompt change.