Write Like It's 1866: LLMs Relearn Telegraphese

TL;DR Highlight

Telling LLMs to write in telegraphese cut output tokens 40–49% with comprehension intact

Who Should Read

A backend developer whose agent memory, scratchpads, and summaries are driving up LLM API output-token costs

Core Mechanics

One sentence cuts output 40–49%
- Tell the model to drop articles and filler, abbreviate, and keep every fact, number, and proper noun verbatim.
- On each provider's own meter, savings were 40.4% for gemma-4-31b, 48.9% for qwen3.8-27b, and 48.4% for GLM-5.3-Flash.
Other models read it without losing accuracy
- Accuracy from cablese records relative to plaintext records was 0.99–1.10, where 1.00 means equal to plaintext.
- This held for four model families never shown an example.
You must ask for lowercase
- Without that, models write cablese in ALL CAPS.
- That costs 14–19 points of savings: gemma 25.0%, qwen 29.8%, GLM 33.9%.
Compress after content is settled
- Answering in cablese while composing scored 0.81 versus plaintext answers.
- Compressing into a record a machine will read scored 1.08–1.09.
Mandatory reasoning backfires

Community Reactions

Opinion that the grading drives the result
- A commenter noted the plaintext baseline is only ~91% because strict phrase matching penalizes correct answers.
- They argued compression-friendly content makes this no proof the phenomenon is emergent.
Experience that the grader is the real variable
- One commenter saw F1 or a judge model swing passage-QA scores by ten points on identical answers.
Speculation that vendors already do this
- A commenter said leaked reasoning in GPT5.6 conversations looks like terse notes, suggesting something similar is already in use.
Opinion that token-saving hacks will fade
- A commenter argued context compaction and AGENTS.md tricks will go away as telegraph-era tricks did.
Complaint about the article's style
- A commenter said the page is so laden with Claude-speak that the facts are hard to find.

How to Apply

Add the cablese instruction to scratchpad summaries
- Today, append the one-sentence lowercase cablese instruction to your summary or memory-store prompt.
- Compare output tokens before and after on the same 10 inputs.
Apply only after content is settled
- Use it for agent handoffs, memory stores, and scratchpads.
- Do not use it while the model is still composing an answer.
Compress with models whose reasoning you can disable
- Pick a compression model that lets you turn reasoning off.
- Drop any model whose writes cost more than plaintext.
Measure savings and accuracy on your own model
- The same instruction gave 40–49% across models, so measure yours.
- Run the benchmark at github.com/Travis42/telegraph-test.

Try It

copy & run

# Paste into Claude or ChatGPT as-is (the article's method turned into a prompt)

Write the document below as a complete record in telegraphese. Drop articles and filler, and abbreviate. Keep every fact, number, and proper noun verbatim. Write in lowercase, not all caps.

Document:
{{문서}}

Next: in a new chat, paste only the cablese result and ask:
{{질문}}
Then compare against the answer from the original document.

Terminology

cableseA compressed writing style from the era when telegrams were billed per word. Like SMS shorthand, it keeps only the key words.

codebookA dictionary that swaps long phrases for single code words. It only renames text already written, so compression is weak.

recovery ratioAccuracy from the compressed record divided by accuracy from the plaintext record. 1.00 means equal to plaintext.

McNemar's testA test on paired results that looks only at questions where exactly one condition was right, to judge whether a difference is luck.

scratchpadA notepad where an agent jots intermediate thoughts.