TL;DR
Telegraph English (TE) is a text-compression format where an LLM rewrites long text into a compact "telegraph" language: it drops filler words, replaces them with ~40 logical symbols (→, ∧, ↑, ∴), and turns every statement into its own line. It doesn't delete words — it rewrites while preserving meaning.
The standard way to compress text is to ask the model to "shorten" or "summarize" it. The problem: the model loses details, breaks logical links and throws out numbers. Ask it to summarize the terms of a contract and you get "the parties agree to fulfill their obligations" instead of concrete deadlines and penalties. The reason: when shortening, the model decides what's "important" — and it often gets exactly the small stuff wrong.
TE solves this differently: the text isn't shortened — it's re-encoded. Each line = one statement. Each link = a specific symbol (→ is causation, ∴ is a conclusion, ∵ is a reason). Numbers always come with units. The upshot: 50% fewer tokens, but 99.1% accuracy on key facts.
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Method Blueprint
Everything happens in a single request to the LLM — the model compresses on its own following the TE grammar:
STEP 1: Load the TE grammar → a system prompt with rules and symbols
STEP 2: Provide the text to compress → the LLM rewrites in 6 passes within one call:
(1) identify concepts
(2) extract statements
(3) mark links with symbols
(4) remove duplicates
(5) verify numbers
(6) verify references
STEP 3: Get the compressed text → each line = one fact, each link = a symbol
STEP 4: Use the compressed text → paste it into later requests instead of the original
Example in Action
Task: You're reading a long deep-dive from The Information or Morning Brew about an investment case (15,000 characters). You want to ask questions, but the text doesn't fit in the context — or it just takes too long to read. You need to compress it without losing details.
Prompt:
You are a text compressor that outputs Telegraph English (TE).
TE GRAMMAR:
SYMBOLS:
= definition/equality
→ causation/consequence
⇒ logical implication
∴ conclusion
∵ because/reason
↑/↓ increase/decrease
∧/∨/¬ and/or/not
≈ approximately
VS comparison (not causation)
RULES:
1. One line = one statement/fact/event
2. All terms in CAPITAL LETTERS (except names and SI units)
3. Numbers always with units: +27.5%, USD10M, 30D
4. References: [AUTHOR:YEAR]
5. TAGS: H1: H2: — sections; CTX: — shared context; PAST: NOW: FUTURE: — time; CONF=0.87 — confidence; DEF: — definition
6. At most 3 symbols in a row on one line
7. You MUST NOT drop information — only re-encode it
CHECK BEFORE OUTPUT:
□ All numbers preserved with units
□ Logical links expressed with symbols, not words
□ No dangling pronouns (not "he" but the entity's name)
□ Each line is one atomic statement
Compress the following text according to the TE rules:
[PASTE TEXT HERE]
Result:
The model returns structured text — headings with H1:/H2: tags, every fact on its own line, numbers preserved exactly, causal links marked with symbols. About half the size. You can paste this text into your next request ("here's the compressed analysis, answer my questions") — and the model will digest it more accurately than a chopped-up, chaotic excerpt.
Why It Works
The LLM's weakness when shortening: When a model "summarizes," it decides on its own what matters. The number 4.8% next to a long paragraph looks insignificant — and gets cut. The condition "no later than 30 calendar days" turns into "30 days." The link "X because Y" turns into "X. Y." — and it's no longer clear whether one caused the other.
The LLM's strength: The model is excellent at following detailed structural instructions and working with formalized patterns. If you say unambiguously "each line = one fact, a number = always with a unit," the model carries it out mechanically and reliably.
How TE uses this: Instead of "decide what's important" — "re-encode by the rules." The grammar doesn't allow numbers to be dropped. The → link can't get lost — it's written explicitly as a symbol. The pronoun "he" is replaced by the entity's name — no dangling references. The model doesn't guess what's important; it keeps everything in compact form.
Control levers:
- Drop the CAPITALS rule → more readable for humans, slightly worse for machine processing
- Adjust CONF=0.87 → add it to the output wherever the source is unreliable — the confidence level is visible at a glance
- Add CTX: → pin the shared context once instead of repeating it on every line
- Add PAST:/NOW:/FUTURE: tags → especially useful for strategy documents and timelines
Prompt Template
You are a text compressor that outputs Telegraph English (TE).
TE SYMBOLS:
= (definition) | → (cause→effect) | ⇒ (logical implication)
∴ (conclusion) | ∵ (because) | ↑/↓ (increase/decrease)
∧/∨/¬ (and/or/not) | ≈ (approximately) | VS (comparison)
RULES:
1. One line = one statement
2. Terms in CAPITALS (except names and units: kg, %, USD)
3. Numbers always with units: +{percent}%, {amount}M, {days}D
4. Tags: H1:/H2: — structure; CTX: — context; PAST:/NOW:/FUTURE: — time
5. Don't drop information — only re-encode
CHECK BEFORE ANSWERING:
□ All numbers preserved with units
□ Links expressed with symbols
□ No pronouns without a link to an entity
□ Each line = one statement
Compress this {document_type}:
{text}
Placeholders:
- {document_type} — article / contract / interview / correspondence / report
- {text} — the text to compress
- {percent}, {amount}, {days} — numeric values from the text, filled in automatically by the compressor
🚀 Quick start — paste into the chat:
Here is a Telegraph English template for compressing text.
Adapt it to my task: {your task}
Ask me questions to understand what to compress.
[paste the template above]
The LLM will ask what exactly to compress and why (to reuse in a chat? for storage? for analysis?) — because the goal determines how aggressively to compress and which tags to add. Then it will adapt the template and compress your material.
Limitations
⚠️ Short texts get longer: Dense text with no fluff (technical formulas, data tables) will grow in tokens under TE compression. The method works on bulky informational texts full of paraphrasing and introductory phrases.
⚠️ Subtle distinctions get lost: TE compressed "no later than 30 calendar days" into
DEADLINE=30D-AFTER-NOTICE— and lost the word "calendar." Wherever legal or contractual precision is critical, check the compressed version by hand.
⚠️ Quality depends on the compressor model: GPT-4o-mini and weaker models apply the grammar less well. Use GPT-4.1 / Claude Sonnet for compression — and then a weaker model can read the compressed text.
⚠️ English only in the original: The grammar was designed for English. On texts in other languages it works with simplifications — symbols and tags carry over, but the CAPITALIZED terms look unusual.
How They Studied It
The researchers took 503 long documents from LongBench-v2 (technical reports, dialogues, multi-document analysis) and sliced them into 1,000-word chunks — yielding 4,081 fragments. Each was compressed two ways: Telegraph English and LLMLingua-2 (the strongest existing method). Then GPT-4.1 generated questions about the original plus "almost right" answer options (e.g., 4.8% vs 4.3%). Five models answered the questions reading either the original or the compressed version.
An interesting detail: the researchers deliberately built an adversarial test (fine_facts — 801 pairs) targeting exactly the details that get lost in compression: precise numbers, conditional qualifiers, edge cases. It's an honest test — they went straight for the weak spot.
The result was surprisingly predictable: the ranking never broke once — original > TE > LLMLingua-50% > LLMLingua-33%. Meanwhile, on small models the TE vs. LLMLingua gap grew to 11 percentage points. The logic: weak models can't reconstruct implicit links from fragments — TE spells them out explicitly with symbols, so the model doesn't have to "fill in" anything.
Original from the Paper
"According to research by Johnson and colleagues (2023), the application of machine
learning techniques to medical diagnostics resulted in a 27.5% increase in early detection
rates while simultaneously reducing false positives by approximately 12% compared to
traditional methods."
↓ After TE compression:
ML → MEDICAL-DIAGNOSTICS: EARLY-DETECTION+27.5% ∧ FALSE-POSITIVE↓12%
[JOHNSON:2023]
Context: The authors show the difference between 68 tokens in the original and 14 tokens in TE — with causation, both numeric claims and the citation fully preserved.
A clinical example from Appendix A:
H1: CLINICAL-TRIAL OUTCOMES
CTX: PHASE-III RANDOMISED CONTROLLED-TRIAL(RCT); N=2400
PRIMARY-ENDPOINT: MORTALITY↓23% VS PLACEBO; p<0.001 [SMITH:2024]
SECONDARY-ENDPOINT: HOSPITALIZATION↓18%; p=0.003
ADVERSE-EVENTS: NAUSEA=12% ∧ HEADACHE=8% ∧ SERIOUS=2.1%
H1: SUBGROUP-ANALYSIS
AGE>65: MORTALITY↓31% (STRONGER-EFFECT)
AGE<65: MORTALITY↓14% (WEAKER-EFFECT)
Adaptations and Extrapolations
💡 Adaptation: compressing a long conversation to carry over into a new chat
The problem: You started a conversation in ChatGPT, it's grown to 50 messages, and you want to move it to a new chat without losing context.
Compress our conversation into Telegraph English.
Rules:
- H1: — topic of the block
- CTX: — the task and conditions (once, at the start)
- DECISION: — decisions made
- OPEN: — open questions
- Each fact, decision, argument — on its own line
- Symbols: → (consequence), ∧ (and), VS (comparison), ↑↓ (increase/decrease)
Conversation to compress:
[paste the chat history]
Paste the result into a new chat with the line: "Here's the compressed context of our previous work in TE format. Pick up from here."
🔧 Technique: add CONF= to rate source reliability
If you're compressing material of varying reliability (articles vs. rumors, facts vs. forecasts):
Add a CONF=0.X tag to every statement:
CONF=1.0 — verified fact
CONF=0.7 — expert opinion
CONF=0.4 — assumption/forecast
CONF=0.2 — unconfirmed data
Result: the compressed text shows at a glance what's fact and what's speculation — no need to reread the original.
🔧 Technique: "TE for prompts" — write in TE style yourself
The principle works in reverse too: write your own prompts in TE style to save tokens and remove ambiguity.
Instead of:
I need you to analyze the competitors of my startup,
which does food delivery in US cities with populations
under 500,000, and focus specifically on price
positioning and delivery speed.
Write:
TASK: COMPETITOR ANALYSIS
CTX: STARTUP=FOOD-DELIVERY; MARKET=US-CITIES<500K
FOCUS: PRICE ∧ DELIVERY-SPEED
Shorter, sharper, and the model doesn't lose a single condition.
Resources
- Paper: Telegraph English: Semantic Prompt Compression via Structured Symbolic Rewriting
- Authors: Mikhail L. Arbuzov, Ziwei Dong, Sisong Bei, Dmitri Kalaev, Alexey Shvets (Palo Alto Networks)
- Dataset: LongBench-v2 (Bai et al., 2024) — arXiv:2412.15204
- Baseline: LLMLingua-2 (Pan et al., ACL 2024)
- Related work: MemGPT (Packer et al., 2023), Chain-of-Thought (Wei et al., 2022)