JuliusBrussee

Current agentic coding setup I don't have anything special going on with my setup but I do like lists of tools to try out. I can't believe it's only been a year since I first wrote about my experiments with coding agents. For side projects and personal stuff, I have a completely wild and constantly changing array of tools. I'll talk about my work setup here since it's more consistent and I do get a decent amount done with it every day. We use GitHub Copilot so this is tailored to that. Coding Harness Generally pi and/or the copilot app. I like working in the terminal and I prefer pi for that but the copilot desktop app has a lot of really nice features. Things like spawning subagents, setting up schedules, managing context are all pretty nice and visual in the app. Terminal Setup I use herdr to manage my pi sessions. My pi sessions can also use herdr to spin up other herdr sessions. I love tmux but the visual aspects of herdr work better for me. I use Ghostty as my terminal emulator. Tools Interacting with agents Plannotator is a great tool and fun to use. I use it with pi to iterate on planning steps and for reviews. The copilot app has a built in review tool that is similar. APM MCP (or whatever is available to interact with your APM) - won't go into specifics about what we use but giving agents read access to your APM and any other observability is so useful for debugging issues but also iterating on a design, identifying performance improvements, looking for error patterns and generally understanding how your app works beyond just the code. Reducing token usage - RTK is an easy to use proxy for cli commands to optimize output for agents. - context-mode optimizes token usage in tool use, so far seems to have the biggest impact for me but I haven't gathered enough data to say for sure yet. - Headroom is a proxy between your harness and LLM api, reduce token usage by stripping out unnecessary context and other optimizations I haven't fully understood yet. I created a wrapper CLI headroom-kit to help me use it with the copilot app and vscode. If you use nix headroom-kit-nix. Skills I find it a little cringe when people share their custom skill stack like they've engineered some amazing thing, but maybe I should try them out more. I find iterating on my own skills for my own work a better experience in general but there are some handy skills out there that have saved me time. Here's what I've found useful so far: - Impeccable for when I want the agent to make a UI nice, because I have no idea how to make a UI nice. - Ponytail, I have my own "prepare-pr" skill that uses this to spin up a code review agent, it seems to do a pretty good job of making the code less yuck. - herdr I mentioned already but it's a neat skill for getting agents to spin up herdr sessions and interact with other agents that way. - gh-stack to tell the agent how to setup a PR stack when trying to break one change down into smaller PRs. (I still often read the code and we still do PRs so this is handy). - grill-with-docs, occasionally use this when I absolutely want to torture myself and think about every possible scenario that could come up for a given problem. I usually throw away the docs afterwards but it has helped a few times to do this process. - show-me to get the agent to explain things visually to me, useful in loads of places including concise PR messages. There are related skills like caveman and bro but I haven't found a need to try them yet. If you use nix, I created a plugin to manage my skills skillset-nix. The problem it solves is making it easier to vendor skills and do some basic security scanning so you don't have to nxp install them. It also gives me a way to manage my own custom skills alongside third-party ones. General tips that save me time - When getting an agent to diagnose something I ask it to use the 5 whys technique. I's annoyingly effective. - Use schedules a lot. If you find you are having to check on things, get the agent to push to you instead. Copilot app support this by you just asking it to schedule something, e.g. "let me know then this PR build passes". - Use static analysis and other deterministic checks more and optimize them for agent use. For example, here's a pre-commit plugin I setup to catch a lot of common issue I see when working with coding agents. The output is text telling the agent what to do instead. - Those last two are thing you do when trying to build "loops" or graphs or whatever they're being called these days. Automating as much of the workflow as possible is not agentic coding specific, we've always done that and it does really help. - Consider making skills for things like priming an agent with broader architectural information. I don't know if a skill is actually the right thing to use but I do a kind of /load-everything-about-that-internal-platform-this-uses where I've documented a bunch of things to load context for the agent specific to certain types of tasks, so don't belong in AGENTS.md.

Trying the caveman token-saving skill without fooling yourself: what to compare, and what to switch off first Source: the README of JuliusBrussee/caveman (Apache-2.0). The numbers below are the maintainer's own, not independently reproduced. The suggested checks below are ways to evaluate it, not results from testing it. What it is. A skill that makes the agent answer in short plain sentences, now bundled with other parts: a compressor for logs and data files, a model router called Auto, and saved helper scripts. Each part has its own switch. What the README says about the main claim. In its own test of ten dev questions, a plain "Answer concisely." line cut output tokens from 6,983 to 4,334. On top of that line, the caveman voice cut about 3% more at the median, and the stronger mode about 35% more (measured on older rules, not re-run). It also says that if your plan charges per request rather than per token, shorter answers may save nothing. A safer way to try it (suggestions). - Compare against a plain "be concise" line in your CLAUDE.md, not against no instruction. That is the baseline the README itself uses. - Start with the skill alone. The README says the skill sends nothing. Add the other parts one at a time, since each has a switch. - Judge it on your own sessions. The bigger claims are about shrinking what the agent reads (logs, JSON, CSV, test output), so they depend on what your sessions read. The README lists HTML as 9.9% worse in one run. - Look at your bill or usage page, not only token counts. Your plan's pricing decides whether fewer tokens means less cost. What leaves your machine, per the README. - The skill alone sends nothing. The compress command sends the file you compress to your model provider. - The CLI sends usage stats by default (commands run, token counts, a random install ID, OS and IP, not prompts or code). DO_NOT_TRACK=1 turns that off. - If you sign in and pick Auto, each request sends your latest ask to Caveman Cloud, plus whatever your agent attaches, such as CLAUDE.md, memory or mentioned files. On the free plan the README says Caveman may keep it to improve routing. That is the part to decide on deliberately. Question. If you've used a token-saver, did you compare it with a plain "be concise" first? What did you find mattered?

Does a token-saving skill beat simply asking for shorter answers? The "caveman" token-saving skill publishes where it loses: "Answer concisely" alone cut output 38% in its own test Caveman (JuliusBrussee/caveman, Apache-2.0) makes the agent answer in short, plain sentences. Its README has a section on where it loses, and the numbers are more useful than the pitch. These are the maintainer's own measurements, not independently reproduced. Ten dev questions on claude-opus-5-5, output tokens: - No instruction: 6,983 - "Answer concisely.": 4,334 - /caveman: 4,119 - /ultracave: 2,693 The README's own reading: "Be concise" is the real baseline. On top of it, /caveman cuts 3% more at the median and /ultracave 35% more (measured on older rules, not yet re-run). The rules cost about 1,160 input tokens, and on a purely per-request plan, fewer output tokens alone would not lower the request charge. Other numbers it lists: - JetBrains paired A/B on 86 real coding tasks: no measurable quality loss (p = 0.82), 8.5% fewer output tokens. - A "read less" feature that shrinks logs, JSON and CSV before the model reads them: 33.2% fewer input tokens across 54 Claude Code runs (18 of 18 answers right). HTML came out 9.9% worse, and the README says that run used an older engine. Privacy, per the README: the skill itself sends nothing. The separate CLI sends usage stats by default (commands run, token counts, a random install ID, OS and IP, not prompts or code), and `DO_NOT_TRACK=1` turns that off. Question: if you use a token-saving skill, did you compare it with a plain "be concise" line in CLAUDE.md first? What did you see?

8- Caveman: Codex, Gemini CLI, OpenCode ve 30 dan daha fazlada yerde çalıştırırsın. Pcnizde , anahtarlarınızla ve Claude Pro/Max girişinizle birlikte çalışıyor. JuliusBrussee/caveman

Neither, just be neutral and use caveman speak. It saves tokens. Being polite consumes extra tokens, and being rude can risk you getting banned. So use caveman speak, save tokens. The best part is, you'll get curt responses as well, without any cringy flowery language that LLMs are known for. In fact, there's even a skill for it: juliusbrussee/caveman

Caveman mode for Claude Code: fewer output tokens, same facts --- title: "Caveman mode for Claude Code: fewer output tokens, same facts" published: true description: "Install the caveman plugin for Claude Code, switch between /caveman, /ultracave and /megacave, read its published numbers honestly, and know when to turn it off." tags: claudecode, ai, llm, productivity canonical_url: mahmadmanzoor.com cover_image: mahmadmanzoor.com --- Claude Code is polite. Ask why a component re-renders and you get a paragraph that restates your question, explains the cause, suggests a fix and offers more help. In a terminal, at the end of a long day, I want the fix and the reason. Nothing else. Caveman is a Claude Code plugin (and a skill for many other agents) that does exactly that. Its tagline is "why many token when few do trick". Under the joke is a careful set of writing rules. This post covers what it is, how to install it, the modes, what its own repository measures, and the cases where I turn it off. Everything below comes from the plugin files installed on my machine (version 3.1.0) and the public repository. Where I quote a number, it is the repository’s number, and I say so. JuliusBrussee/caveman: The caveman repository: the skill, the Claude Code plugin, the optional proxy CLI, and the benchmarks behind its numbers. Apache-2.0. What it actually changes The core is one skill file. It tells the model to answer first and then give the reason and the next step, to drop greetings, hedging and recaps, and to use short words. Articles like "a" and "the" are optional. Words that carry meaning are not: the rules say never to drop not, never, no, only or except, and to keep numbers and units exact. The rule I like most is "payload verbatim". Code blocks, commands, file paths, API names and error messages are copied character for character. Only the prose around them gets shorter. The README shows the same React answer three ways: !Bar chart of one React fix written four ways: normal agent 63 tokens, /caveman 20 tokens, /ultracave 14 tokens, /megacave 13 tokens One answer, four voices. Example and token counts from the caveman README (tiktoken o200k). normal reply The reason your React component is re-rendering is likely because you're creating a new object reference on each render cycle. When you pass an inline object as a prop, React's shallow comparison sees it as a different object every time, which triggers a re-render. I'd recommend using useMemo to memoize the object. /caveman reply New object ref each render, so React re-renders. Wrap the prop in `useMemo`. The facts are the same: a new reference each render, React sees a change, useMemo fixes it. What is gone is the ceremony. The skill also says caveman is "a voice, not broken grammar". There is no "me think" and no caveman prefix, and if the terse version is not actually shorter, plain wins. Why fewer output tokens matter !Two boxes: what the agent reads (system prompt, tools, files, logs, conversation), which the skill does not shrink, and what the agent says (prose in replies, code and commands exact), which /caveman shortens; the rules ride along as extra input on every call The skill works on the second box only. There are three reasons, and they are not equally strong. Cost. Output tokens are billed, and on most model price lists they cost several times more per token than input. The skill file puts it simply: every output token is billed and read, so filler costs twice. Speed. A model writes its answer one token at a time. A shorter answer finishes sooner. In an agent session with many turns, that adds up to less waiting. Reading. This is the one I feel most. I read every reply in a terminal. An answer that starts with the cause, not with "Great question", saves my attention, which is the scarcer resource. The important limit: the skill only shortens what the agent says. It does nothing to what the agent reads, which in a coding session is usually much bigger: the system prompt, tool definitions, files, test output and the conversation. The repository’s own Honest Numbers page lists the skill’s input reduction as 0%. For input, the project has a separate, optional proxy CLI, which I cover near the end. Install it in Claude Code 1. Add the marketplace and install the plugin This is the Claude Code route from the README and INSTALL.md. The plugin starts caveman automatically in every session. claude plugin marketplace add JuliusBrussee/caveman && claude plugin install caveman@caveman 2. Or do the same from inside a session The public README lists the slash-command form too. /plugin marketplace add JuliusBrussee/caveman /plugin install caveman@caveman 3. Restart and check the flag file Hooks are read once at session start. INSTALL.md says the active mode is mirrored in a small flag file. cat "${CLAUDE_CONFIG_DIR:-$HOME/.claude}/.caveman-active" # expected output: caveman 4. Update with the full name The plugin and its marketplace are both called caveman, so the short name fails with "Plugin caveman not found". Use plugin@marketplace. claude plugin update caveman@caveman If you only want the skill, without hooks, npx skills add JuliusBrussee/caveman -g installs it for Claude Code and other agents. Then type /caveman to start it. One version note: my machine runs 3.1.0, while the public README now pins its install scripts to v4.0.1. The plugin commands above are the same in both. Modes and levels !Three modes: /caveman is the default voice, /ultracave is fragments with each fact once, /megacave is Classical Chinese; old level names lite and full map to caveman, ultra to ultracave, wenyan-* to megacave Version 3.1 has three modes. Older level names still resolve. There are three modes. /caveman is the default voice described above. /ultracave (alias /caveman ultra) strips grammar further: fragments, one word when one word is enough, each fact stated once. /megacave (alias /caveman wenyan) answers in Classical Chinese, with technical terms left in their original script; its own skill file says the token savings are small and noisy and must never be claimed. Older releases talked about levels: lite, full, ultra and wenyan. The 3.1 config code still accepts those names and maps lite and full to caveman, ultra to ultracave, and any wenyan variant to megacave. A mode sticks for the whole session until you change it. - /caveman status reports the current mode and changes nothing. - /caveman off, "stop caveman" or "normal mode" switches it off. - /caveman-help prints the reference card. You can also change what a new session starts in. The help card lists an environment variable, which wins, and a user config file. Setting the default to manual means sessions start off and /caveman turns it on, which is how I would roll it out to a team that has not asked for it. terminal # highest priority export CAVEMAN_DEFAULT_MODE=ultracave ~/.config/caveman/config.json { "defaultMode": "manual" } How the plugin wires itself in The plugin is small. Its manifest registers two hooks, and a third script runs when you ask for stats. !SessionStart runs caveman-activate.js, which adds the active mode’s rules to the session; UserPromptSubmit runs caveman-mode-tracker.js, which spots /caveman ultra or off and saves the mode per session; /caveman-stats runs caveman-stats.js, which prints recorded usage from the session log Two hooks and a stats script. Each Claude Code window keeps its own mode. Each window keeps its own mode in a file under .caveman-sessions/ in your Claude config folder, so two sessions can run different modes. /caveman-stats reads the session log and prints the tokens actually recorded. It reports savings as unknown, because a session has no measured twin without caveman to compare against. I like that honesty more than a made-up percentage. The plugin also ships helpers I use more than I expected: /caveman-commit for one-line Conventional Commit messages, /caveman-review for one-line review findings, and /caveman-compress <file> for shrinking memory files like CLAUDE.md while keeping headings, code blocks and paths. The repository reports about 46% smaller files across five fixtures for that last one. What the numbers say, and what they do not !Output tokens over ten developer questions: no instruction 6,983, Answer concisely 4,334, /caveman 4,119, /ultracave 2,693 The repository’s own eval on claude-opus-5-5. Ten questions, one run, output length only. The README’s eval asks ten developer questions on claude-opus-5-5. With no instruction, the answers totalled 6,983 output tokens. With "Answer concisely." they totalled 4,334. /caveman gave 4,119 and /ultracave gave 2,693. The repository makes the point itself: new models already know how to be concise, so that one line is the fair baseline. Against it, the Honest Numbers page reports caveman 3% shorter, ultracave 35% and megacave 9% at the median, from ten questions in a single run, measuring length only. The README also cites outside work: a JetBrains A/B over 86 coding tasks that found 8.5% fewer output tokens and no measurable quality loss, an Adobe Research paper reporting 1.4 to 2.4 times lower cost for caveman-style output across eight models, and an Elastic adaptation with 63.6% fewer response tokens. Those are their numbers, in their setups. The honest summary for plain /caveman on a modern model is: a little shorter than asking for concise answers, and noticeably easier to scan. When terse output hurts Caveman has a cost side, and the project documents it better than most tools do. - The rules are extra input on every call. The 3.1 README estimates about 1,000 input tokens per call and the current one about 1,160; the Honest Numbers page says the real cost depends on caching and was not measured. In short coding Q&A, one user measured a net loss (issue #145). - Per-request billing gets nothing. If your plan charges per request rather than per token, as GitHub Copilot premium requests do, a shorter answer costs the same. - Learning needs full sentences. When I am new to a codebase or a library, I want the explanation, not the verdict. Fragments assume you already share the context. - Two readings of one fragment. "Migration drop column backup first" can be read two ways. The skill’s own fix is "Back up first. Then run migration: it drops the column." Ambiguity is worth more tokens than it costs. !Cases where caveman writes in full sentences: security warnings, irreversible actions, steps where order matters, a confused user, and anything persisted such as code, commits, docs and PRs Built-in exceptions. After them, it goes back to terse replies. To its credit, the skill switches to plain prose by itself for the riskiest cases: security warnings, irreversible actions, step sequences that a fragment could scramble, a user who seems confused, and anything written outside the chat, such as code comments, commits, docs and pull requests. Those exceptions are the reason I trust it enough to leave on. A note on the proxy The bigger savings the README advertises come from a separate tool: the @caveman-ai/cli proxy, which compresses what the agent reads, such as logs, JSON and test output. It is a different trade-off. INSTALL.md says that routing Claude Code through it sets ANTHROPIC_BASE_URL in your settings, which makes Claude Code Remote Control unavailable while routing is on, and the README says the CLI sends usage statistics by default (caveman telemetry off or DO_NOT_TRACK=1 turns that off). Read both before you install it. The skill alone runs locally and sends nothing. How I use it I keep /caveman on for routine coding sessions, switch to /ultracave for quick yes-or-no checks, and say "normal mode" when I am learning something or need a careful explanation for someone else. If you try it, run the same task with and without it and compare your provider’s usage page. That is the test the project itself recommends, and it is the only number that counts for your workload. --- Originally published on my site. More posts and projects at mahmadmanzoor.com.

Caveman make AI agent talk less, code stay exact, input tokens drop about 33%. 🔗 JuliusBrussee/caveman Credits: JuliusBrussee/caveman #caveman #go #claudecode

「compress-for-llm」 --- name: compress-for-llm description: Compress text to minimal-TOKEN LLM-readable form — lossless, self-decodable. Use when compressing text whose READER is an LLM (agent docs, prompts, notes), NOT human-facing prose. --- # Compress for an LLM reader (¬human) TASK ≜ verbose human text → 最小 **token** 列。reader = 同等能力 LLM ∴ 単一言語文法·敬語·filler 棄。 ## directive - **omnilingual(実測で絞済 2026-07-22)**: ∀ 字体/語 横断 → 概念毎に **最短 tokenize** される表現を 選ぶ。実効 palette = **漢字・中文複合語・文言機能語・English(電報体)・ASCII 記号**のみ。 **英文原文・自作の定型指示文 ⇒ 英語電報体(骨格)が最強**(BPE = 英語 first・"Proc failed => abort."=5tok。 本 skill 自文の英中ごたまぜ再圧縮 = 中文混成 474→401tok 実測 2026-07-22・行別に英/中候補→全文 diff で勝者採用・ 勝ち残り中文行 = `path/CLI/enum/error literal=勿译` のみ)・ニュアンスを失えない日文原文 ⇒ CJK 混成。 独複合語(Geschwindigkeitsbegrenzung=6 vs 速度制限=3)・キリル(запрещено=5)・Arabic/Sanskrit/ 𓂀 等の珍字体 = **実測で敗北** ∴ 「全書記系」は幻想 — palette は tokenizer 学習頻度が決める。 - **search order(候補の探し方・出力文法の強制ではない)**: cut → code-like structure(`key:value;`・`f(x)`・`{a,b}`・`x:=…`) → ASCII operator(`->` `=>` `,` `<=` `not`)。math 記号(⇒ ∴ ∧ ¬ ∈ ≜ …)は **full-string diff で勝った時だけ** (実測 2026-09-25: ASCII 版が 12/12 セルで安い・勝ったのは Σ のみ = §MiL)。先に**意味の同値**を確かめてから token を比べる。 - **cut**: 重複 · 自明語 · 言い換え · 修辞的hedge · 前置き。`[要検証]` / `UNMEASURED` / 条件・例外・owner gate・fail-closed強度は**情報本体**∴削除/弱化禁止。 ## ⚠️ token 節約を理由に CJK 圧縮を選ばない(当 repo 実文で実測 2026-09-01・[[1182]] §3-1) `tiktoken o200k_base`・当 repo canon の実文 2 例を 3 文体で測った(**当方実測 CONFIRMED**): | 例 | 当 repo の圧縮体 | 平文 日本語 | **English 電文体** | |---|---|---|---| | fence 根拠 | 41 tok | 55 tok | **17 tok** | | 委譲規 | 23 tok | 35 tok | **15 tok** | ⇒ ① **圧縮体 > 平文日本語**(∴ 圧縮自体は効いている — 「文言文は逆効果」という外部実測 〔`codex-tokens-compress`: 文字 −34% ∧ token +6%〕は**当 repo では再現しない**)。 ⇒ ② **但し英語電文体が 1.5–2.4 倍有利** ∴ 本 skill 冒頭の「英文原文・自作の定型指示文 ⇒ 英語電報体が最強」を**当 repo の実文で追認**した。 ⇒ **∴ 規: token 削減が目的なら英語電文体を選ぶ。CJK 圧縮を選ぶ理由は「ニュアンス保持」 (日文原文の含意・owner 語彙・法/規の逐語)だけであり、token 節約ではない。** ⚠️ **文字数で自己評価しない**(低頻度漢字は BPE で分割され得る)— 判定は必ず tokenizer で測る。 ∧ 当 repo の追跡済 tier 集計は B/tok ~3.1 ∴ **tier 単位の肥大検知に byte を使える**。 個別書換では代理に使わず、上記どおり tokenizer で前後を測る。 ## language substitution findings - 日文手順→中文は、助詞/活用を削り複合語を短くできる場合が多い。簡体/繁体/地域語を固定せず、 文脈ごとに短く明確な言い回しを選ぶ。ただし**文字数≠token数**; target tokenizerで比較できなければ 「短い可能性」とだけ扱い、珍字1字化を目的にしない。 - path / CLI / status / enum / schema key / error literalは翻訳しない。否定(`禁止/不得/≠`)・順序・ owner境界・fail-closed条件は省略しない。 - **reader tier で分ける**(owner 裁定 2026-07-11・Gemini 復号実証 = [[0469]]): 同等能力 LLM 向け (本 skill の既定 scope)= **容赦無 palette 維持**(閩南語 毋通/著愛・chữ Nôm 等も同 tier では復号実証済)。 共有 skill / SOP(弱 model も読む)= AGENTS 弱model耐性が優先 ∴ 中文の普通語+明示動詞・ 一字略語/古語/言語混合で復号に推論を要する形は棄却。 - 圧縮後に fresh reader が `{trigger, steps, output, prohibition}` を原文無しで復元できるか確認する。 **Fresh reader**: `tools/probe/fresh_reader_quiz.py`; frozen fixtures, English literal-copy contract, strict gold / UNKNOWN / absent-fact fabrication counts. Set `--wall-time-cap`; timeout/incomplete = `UNMEASURED`. `--min-correct` explicitly enables a gate; report exit 0 alone proves no floor. Models/command = `.claude/skills/verification-tool-index/index-8-llm.md`. Revisit failed wording; one reader's pass does not establish stronger-reader recovery or other-lane quality. ## prior art(2026-07-22 一次照合済・全 URL fetch 検証済) - **Caveman**(`https://github.com/JuliusBrussee/caveman`)= **出力**側を電報体英語化する skill (自己bench 10 promptのoutput token平均 −65%主張・補正 doc 有・文言文 `wenyan` mode 有)。 本 skill との差分 = 入力/文書側 ∧ 全文 diff 実測手順。 - **ai.rs 検証**(`https://ai.rs/ai-developer/classical-chinese-agent-memory-compression`)= 文言文単独の「−28%」 主張を実測補正(o200k = −13%・cl100k = 悪化)→ **「文言文 + 英語キーの hybrid が最適」= 本 skill の 英中ごたまぜと独立同着** — 結論の外部再現。文字数≠token の罠の実例でもある。 - 学術: LLMLingua 系(`https://arxiv.org/abs/2310.05736` 等)= perplexity 間引き系 = 別 lane (手書き圧縮と競合しない)。Petrov et al.(`https://arxiv.org/abs/2305.15425`)= 同内容でも tokenizer・言語依存で token 差最大15×。**普遍的な英語優位は示さない** ∴ tokenizer別実測が正典。 ## safe symbol palette - **ASCII 同等物(`=>` `,` `not` `in` `<=` `:=`)を先に試す**。以下は diff で勝った時のみ。 - logic: `∀ ∃ ∄ ∧ ∨ ¬ ⇒ ⇔ ∵ ∴ ≜ ≠`。集合/写像: `∅ ∈ ∉ ⊂ ⊆ ∩ ∪ |S| f:A→B`。 - 比較/範囲/集計: `< ≤ > ≥ [a,b) Δ Σ`。金融量は初出定義後のみ `P_t q ΔP gross net`。 - `⊥`等の多義記号、未定義の私製glyph、権限/禁止を弱める記号置換は不可。 - **critical negation = 動詞 + 対象/範囲で書く**(`Do not re-add spread`・`--wait-lock does not permit concurrent broker access`)。 `¬`/`not` を operator・断片の前に付けない(`not +spread`・`¬⇒` は 4 回極性反転 = §MiL)。 - token削減はtarget tokenizerで実測。記号が複数token又はfresh readerの復元を悪化させるなら語へ戻す。 - **実測 tier(2026-07-22・o200k proxy・単独入力計測・Claude tokenizer 未計測 [要検証])**: 1tok 勝者 = 単漢字(故無済未禁必要警増減全)・中文複合語(没有/不可/因为)・**文言機能語(勿莫弗非唯止須宜)**・ ASCII(`=> := -> because must done`)・`⇒ → ¬ ∀ ↑ ↓`。**2-3tok = 集合系記号(∴ ∵ ∅ ∈ ≜ ≠)・ 絵文字(⚠️=3・❌=2)・方言字(冇唔毋 = 粤語 — 密度高くても学習頻度低 = 珍字罠)**。日文行の 中文(簡体+文言調)訳 = 12 行中 11 勝(−10〜15%・skill 全文で 449→428tok 実測)∴ 中文化 = 有効な第二 pass。 - **⚠️ 単独計測 ≠ 文中効果**(BPE は空白/句読点と結合): 実測で「。故、」= 「 ∴ 」= 文中 9tok 同点 (単独では故1/∴2)∴ **最終判定 = 置換前後の全文 diff で測る**。固定辞書扱い禁・手順(測って選ぶ)が主役。 集合記号は「復元の曖昧性を消す」価値がある時のみ。意味非同値(∈≠是・∅≠無)は critical 文脈で語を残す。 ## MiL 検証(2026-09-25・reader = 同等能力 LLM tier のみ・共有 skill/SOP には適用しない) 対象 = @AM09_21 の CC0 圧縮言語 `MiL;C(m):=argmin_tok{s:⟦s⟧=m};R:=sym>logic>abbr>alias>>>EN;del(recoverable);IF:=cond;@:=ext;¬infer;amb→qmin`。 実測: 当 repo 4 文 × o200k/cl100k/Qwen2.5 + qwen3.5:4b 復元 quiz(2 文 × 4 変種・各 1 run・key 事前固定)。 **n 小 ∧ reader 1 model ∴ 以下は傾向であって普遍則ではない**。 - 採る: C(m)(意味固定で token 最小)= 本 skill と同じ目標。code-like structure は b(英語電報体)より安く・復元同等(10/13)。 - 採らない: R の記号優先(実測で逆)・`@` の多義(定義 ext ≠ 使用例 at)。 - **recoverable := 与えられた文面だけから一意に再構成できるもの**(書き手が「推測できる」と思うものは含まない)。 verbatim 不変物(id・owner gate・権限を絞る語・例外・単位・数値・literal・`[要検証]`)を上書きしない。 aggressive del(recoverable) は 4/13 ∧ 禁止権限を「許可」と読ませた。 - **brief 冒頭の reader 契約(prompt/brief 限定・doc/memory には付けない)**: `Do not invent missing facts. If ambiguity blocks action, ask the smallest resolving question; headless: mark UNKNOWN, stop dependent action, continue independent work.` (計算・論理的演繹は禁じない。不在事実 probe は 8/8 UNKNOWN だが契約なし対照は未測定) - **tokenizer 差は target で測る**: Qwen 系は数字を 1 桁ずつ割る(実測 1 例 +24% vs o200k)。補正率を固定しない・数値は削らない。 ## hard constraints (¬破) 1. **lossless**: ∀ fact + detail が出力から復元可能。density は情報を消さない。 2. **self-decodable**: codebook 無しで**想定 reader tier**(同等能力 LLM 群)が復元できる。density ↔ 復元性 = 実 trade ∴「fresh な reader が復元不能になる手前」で止める。「任意の LLM」は検証不能 ∴ 掲げない。 3. **目的 = 最小 token ≠ 最小語**: 珍字は 1 概念でも多 BPE token になり得る ∴ exotic = 自動的に短 ¬。 復元性を保ちつつ **最短 tokenize** を選ぶ(多くは dense な常用字 + 記号; exotic glyph は真に勝つ時のみ)。 ## reader tier × palette(owner 裁定 2026-09-02「LLM 間は容赦不要」・AGENTS.md 鉄則が正典) | reader | 容赦 | 既定 palette | 備考 | |---|---|---|---| | owner(人) | 対象外 | 日本語散文 | 本 skill の scope 外 | | 同等能力 LLM(Codex/Claude/Gemini・自分の次 session・memory) | **無** | **English 電文体**(token 最安)・CJK 混成は nuance 保持時のみ | 敬語/前置き/hedge/保護 boilerplate ∅ | | 中華圏 model(Kimi/Qwen/DeepSeek・roster 外・将来) | 無 | **English 電文(最安)> 简体普通话(nuance 時)> 文言 hybrid** — 実測 2026-09-02([[1202]] §3・Qwen2.5/DeepSeek-V3/Kimi-K2 tokenizer): en 210–215 / 简体 228–240 / 文言 250–282 / 繁體 247–290 / 粤語 266–314 tok(同一 7 行)| 文言が简体に負けるのは珍字 BPE 分割(§token の罠を中華圏 tokenizer でも追認)。繁體は简体比 +9〜27%。再測定 = 対象 tokenizer を指定した scratch script を用意する(実測時の `tok_compare.py` は repo に無い) | | その他言語圏 model | 無 | 相手の学習頻度が高い言語 | **固定しない・選択肢を常に残す**(owner 明示要求) | | 弱 model(共有 skill/SOP の読者) | 有(弱 model 耐性) | 普通語 + 明示動詞 + checklist | AGENTS 冒頭の規が優先 ∴ 本表の「容赦無」は**会話**の規で配布物には及ばない | 不変(全 tier): cmd block・path/flag・数値・重大度 tag・引用・`[要検証]`/`UNMEASURED`・marker・schema required。 ## memory 棚卸し pass の実測知見(2026-09-02・44 file・o200k・reader = 自分の次 session) - **削減の主因 = cut(重複・言い換え・前置き・「Why」の再説明)であって記号/CJK 化ではない**(2026-09-02: 42 file −34%・B/tok 3.0 → 2.9 = palette ほぼ不変)。既に密な note では **prune(repo 記録済 → pointer 1 行)+ 統合(同主題 2 file → 1)が先**(`memory-pruning-freshness` と同順)。統合で消せるのは「撤回済方針」型(教訓だけを親へ 3 行で吸収)。 - **description / 索引 = recall の hook ∴ 1 行で「何の話か」だけ**(2026-09-25: `MEMORY.md` −55% が最大の lever)。索引に結論・数値を写すと、本文だけ訂正された時に古い結論を運ぶ。 - 英語原文の JP 電文化は token 減が小さい。owner 語彙・引用・日付付き実害を含む note は CJK 混成で nuance を保つ(判断基準 = §token)。 - **正確さ > 圧縮率**: 鮮度照合で訂正が増えた file は伸びる。並列 subagent に file 群を割っても、各訂正は当方が source を 1 件ずつ照合する(反証済の主張が 3 群で 15 件以上出た = 照合の方が価値が大きい)。 - 書込経路の罠: Bash heredoc は 1 command ~8 KB 超で `unexpected EOF while looking for matching` → 長文は Write tool。 - **測る**: `uv run --with tiktoken python <scratch>/tokcount.py <dir>`(o200k・file 別 tok/B/B-per-tok)を前後で撃つ。B/tok 不変で tok 減 = cut が効いた・B/tok が動いた = palette が効いた(英語電文化では上がる: 2026-09-25 実測 2.9 → 3.5)。 ## output 圧縮結果**のみ**(前置き·説明 ∅)。入力の構造(見出/箇条)は losslessly 潰せる時のみ潰す。 ## code docstring / comment(owner 裁定 2026-09-06・規 = `.claude/rules/src-comments.md`・実測 = [[D01295]]) - 対象 = 根拠 prose(why / 契約 / 罠)だけ。**逐語のまま** = id(ADR/review/[[NNNN]])・数値・path/flag・外部仕様引用・否定・例外・`[要検証]`・governance が読む形(`Exit codes:`・`EXERCISE`・`Usage:`)。 - 型 = 英語電報体(`Fail closed: None fixture = caller opted out of venue check (ADR-0286).`)。同 file 内の反復説明は 2 回目から id のみ。 - 削る ≠ 圧縮: banner / 手順番号 / 名前 = docstring / 履歴叙述は cut・根拠は compress。cp932 可の文字だけ(`test_tool_console_encoding`)。

I'M WRITER, TECHNICAL WRITER, HIGHLY TECHNICAL PERSON EXECUTIVE SPACE. COMMUNICATION HAS ALWAYS ONE STRONG POINTS, BLATHERING AI DOES DRIVES ME DISTRACTION. 'S IMPOSSIBLE REALLY TAILOR OUTPUT LLMS VERY MUCH. I VALUE INFORMATION DENSE, DIRECT WRITING WHEN COMES TECHNICAL THINGS THEY'RE ALWAYS EMBELLISHING GOING OFF SIDE QUESTS 'S OBNOXIOUS. UGH. juliusbrussee/caveman

개발자들이 먼저 깔았다..꼭 알아야 할 토큰 절약 도구 깃허브 TOP 7(깃허브 ★51만) AI 코딩할 때 토큰이 새는 네 군데(쓰는 양·읽는 양·찾아 헤맴·내 설정)와, 그걸 막으려고 개발자들이 만든 깃허브 도구 7개를 정리했습니다. 00:00 오프닝 00:28 포니테일 01:47 케이브맨 02:54 RTK 04:15 헤드룸 05:27 컨텍스트 모드 06:31 세레나 07:39 토큰 옵티마이저 08:39 마무리 📌 영상에 나온 도구 링크 1. 포니테일 (Ponytail) DietrichGebert/ponytail 2. 케이브맨 (Caveman) JuliusBrussee/caveman 3. RTK rtk-ai/rtk 4. 헤드룸 (Headroom) headroomlabs-ai/headroom 5. 컨텍스트 모드 (Context Mode) mksglu/context-mode 6. 세레나 (Serena) oraios/serena 7. 토큰 옵티마이저 (Token Optimizer) alexgreensh/token-optimizer [en-US] Developers' Must-Have Token Saving Tools: Top 7 GitHub Repos (517K Stars) I have summarized the four places where tokens leak when coding with AI (writing volume, reading volume, searching, and my settings), and seven GitHub tools created by developers to prevent them. I have compiled a list of these tools. 00:00 Opening 00:28 Ponytail 01:47 Caveman 02:54 RTK 04:15 Headroom 05:27 Context Mode 06:31 Serena 07:39 Token Optimizer 08:39 Wrap-up 📌 Links to tools featured in the video 1. Ponytail DietrichGebert/ponytail 2. Caveman JuliusBrussee/caveman 3. RTK rtk-ai/rtk 4. Headroom headroomlabs-ai/headroom 5. Context Mode mksglu/context-mode 6. Serena (Serena) oraios/serena 7. Token Optimizer alexgreensh/token-optimizer [id] Alat Penghemat Token Wajib Punya Pilihan Pengembang (Total 517 Ribu Bintang di GitHub) TOP 7 Saya telah merangkum empat tempat di mana token bocor saat melakukan pengkodean dengan AI (volume penulisan, volume pembacaan, pencarian, dan pengaturan saya), dan tujuh alat GitHub yang dibuat oleh pengembang untuk mencegahnya. Saya telah menyusun daftar alat-alat tersebut. 00:00 Pembukaan 00:28 Ponytail 01:47 Caveman 02:54 RTK 04:15 Headroom 05:27 Context Mode 06:31 Serena 07:39 Token Optimizer 08:39 Penutup 📌 Tautan ke alat yang ditampilkan dalam video 1. Ponytail DietrichGebert/ponytail 2. Caveman JuliusBrussee/caveman 3. RTK rtk-ai/rtk 4. Headroom headroomlabs-ai/headroom 5. Context Mode mksglu/context-mode 6. Serena (Serena) oraios/serena 7. Token Pengoptimal alexgreensh/token-optimizer