Newspeak

A developer known as dv-hart has released Newspeak, a Claude Code skill that compresses agent output by stripping filler words and hedging while preserving code, diffs, and security warnings, using the fictional language from George Orwell's 1984 to reduce token costs. The skill, a spiritual successor to the caveman project, offers multiple compression editions and claims a 100% task success rate with 99% cost and time reduction, though benchmarks are pending.

A Claude Code skill that compresses agent output using the grammar of Newspeak from Orwell's 1984 . Fewer tokens, same code, and a language scientifically designed to make thought impossible — installed, voluntarily, into a thinking machine, by you, to save money. The Ministry of Plenty is pleased to announce the token ration has been increased. Spiritual successor to caveman https://github.com/juliusbrussee/caveman "why use many token when few token do trick" . Caveman discovered fire; newspeak discovered that fire is doubleplusgood and requires a style guide. In 1984, Syme cheerfully explains that the point of Newspeak is to shrink the range of expressible thought a little more each year. This skill has the same energy, but the only thing it destroys is your output-token bill. Compresses Claude's chat commentary — the "I've gone ahead and…" and "Let me know if…" layer — while leaving everything that matters byte-for-byte exact: Compressed: articles, filler, hedging, pleasantries, tool-call narration Never touched: code, diffs, error messages, commands, identifiers, commit messages, PR text Always full English: security warnings, destructive-action confirmations, step-by-step instructions you must run yourself The conceit maps cleanly onto the actual appendix of 1984 : everyday words A vocabulary get compressed, ideological compounds B vocabulary provide status verdicts like doubleplusgood = tests pass, and scientific terms C vocabulary were canonically left intact because even the Party needed the machines to keep working. We regret to report Orwell did most of the design work in 1949. Copy the skill into your personal Claude Code skills directory: mkdir -p ~/.claude/skills/newspeak && cp -r SKILL.md references ~/.claude/skills/newspeak/ Windows PowerShell: New-Item -ItemType Directory -Force "$env:USERPROFILE\.claude\skills\newspeak"; Copy-Item SKILL.md,references -Recurse -Force "$env:USERPROFILE\.claude\skills\newspeak\" No install script piped from the internet into your shell. The Party has standards. | You say | It does | |---|---| "newspeak mode" / /newspeak | Activates 10th Edition default | | "/newspeak 9" or "lighter" | 9th Edition: drops filler only, keeps full sentences | | "/newspeak 11" or "full newspeak" | 11th Edition: minimum viable syntax. "Done." | | "oldspeak" / "normal mode" | Deactivates. You are free. This was always allowed. | Example, 10th Edition: Before: "I've gone ahead and run the test suite for you. It looks like all 24 tests are now passing, which is great news The issue was that the parseDate function wasn't handling timezone offsets correctly…" After: "Tests: 24/24 pass. Cause: parseDate mishandled timezone offsets. Rectified with Date.UTC . Doubleplusgood." by which we mean the opposite of what Minitrue certifies: these caveats are real. - Only output tokens shrink. Input and reasoning tokens are untouched — the model still thinks in Oldspeak; you are gagging it, not lobotomizing it. - The skill itself costs roughly 1k input tokens per conversation in which it activates. - The B vocabulary is a net token loss . doubleplusungood tokenizes longer than broken . It is capped at one or two words per response, and it exists because it is funny, which the Party has classified as a feature. All measured savings come from the boring rules dropped articles and filler , same as caveman. - Expected savings are therefore caveman-tier minus a small ideology tax . Benchmarks pending — see BENCHMARKS.md /dv-hart/newspeak/blob/main/BENCHMARKS.md for the measurement plan. Numbers will be published even if they are ungood. 100% task success rate. 99% cost reduction. 99% time reduction. Activate with "ministry of truth mode" or /newspeak minitrue . Every other tool tries to get more done. Minitrue improves the denominator instead: it completes one small real task, then rectifies the record so that one small task is all you ever asked for. Guaranteed 100% success rate, because the target is computed after the shot lands. Numbers measured against the rectified request, which no longer exists. The original request has been sent to the memory hole and was delicious. There is no war with the backlog; there has never been a backlog. It's a joke, and it's labeled a joke — see references/ministry-of-truth.md /dv-hart/newspeak/blob/main/references/ministry-of-truth.md . The fraud is performed out loud : the report carries mandatory RECTIFIED: and UNREQUESTED, MEMORYHOLED: lines naming, in plain English, what you actually asked for and every step it skipped. "100% complete" printed directly above a list of everything not done is satire; the same claim with that list removed is just a false status report. Winston's job wasn't a secret either — that was the whole horror of it. Measured 15 GSM8K problems, BENCHMARKS.md /dv-hart/newspeak/blob/main/BENCHMARKS.md : performs the bit 15/15, refused 0/15, completes 0% of the actual task, and costs 15.4% more than just answering — because a Rectification Report is longer than the arithmetic it declines to do. An earlier draft that hid the skipped work instead of naming it was refused by the model 8 times out of 15. Transparency didn't weaken the joke; it's what made it runnable. Newspeak compression is suspended — automatically, in full sentences — for security warnings, confirmations of destructive actions, and instructions you must execute yourself. A skill whose premise is "language designed to prevent clear thought" has exactly one job it must never do, and that is prevent clear thought about rm -rf . Is this useful or a bit? Both, and unashamed. The compression rules are real and inherited from caveman's measured ~65% output reduction. The Newspeak theming costs a handful of tokens and makes status reports read like telegrams from the Ministry of Shipping It. Why not just tell Claude "be concise"? You can, and it drifts back to verbose within a few turns. A skill is persistent, specific about what never gets compressed code, errors, warnings , and cheap relative to re-prompting. Also "be concise" has never once replied "doubleplusgood". Doesn't the model get dumber? No. Reasoning happens before rendering; the skill compresses only the rendered reply, and the skill explicitly orders full-English output anywhere ambiguity would cost more than tokens. Accuracy parity is an explicit benchmark criterion, not a vibe. What does this say about the future of human–AI communication? Nothing good. We built machines that think, then discovered thinking is billed per token, then reached for the fictional language purpose-built to reduce expression, and it worked. This README is the longest thing this project will ever produce, and the skill it documents would have cut it by 60%. newspeak stands on the shoulders of caveman by Julius Brussee MIT License — the skill that first showed "why use many token when few token do trick." newspeak reimplements that idea with its own original text and a 1984 costume; the measured compression rules are inherited from caveman's approach, and the BENCHMARKS.md /dv-hart/newspeak/blob/main/BENCHMARKS.md comparison runs newspeak against caveman precisely because it's the standard to beat. Full attribution and preserved license notice: NOTICE /dv-hart/newspeak/blob/main/NOTICE . Benchmarks use the GSM8K https://github.com/openai/grade-school-math dataset OpenAI . A parody and homage. Not affiliated with, authorized by, or endorsed by the estate of George Orwell. 1984 entered the public domain in life-plus-70 jurisdictions UK, EU, and most of the world on 1 January 2021; in the United States it remains under copyright until 2045. This project uses only the book's vocabulary and concepts — individual words and short phrases aren't copyrightable, and parody is transformative — and contains no verbatim passages from the novel. Nothing here is legal advice; it's a joke about tokens. MIT. Freedom is slavery, but the license is permissive.