[{"data":1,"prerenderedAt":137},["ShallowReactive",2],{"post:cutting-claude-code-tokens-with-caveman-mode":3,"posts":109},{"slug":4,"title":5,"date":6,"dateNice":7,"dateShort":8,"lastEdited":9,"excerpt":10,"tags":11,"readTime":15,"emoji":16,"cover":17,"body":18},"cutting-claude-code-tokens-with-caveman-mode","Cutting Claude Code tokens by ~75% with caveman mode","2026-05-14","May 14, 2026","May 14","2026-05-16","An ultra-compressed prompt style that strips articles, filler, and pleasantries from Claude Code without losing technical substance. Here's what it costs, what it saves, and where it bites.",[12,13,14],"claude-code","tooling","notes",6,"🪨",null,[19,22,24,27,29,31,35,37,39,41,45,47,49,51,54,56,58,60,62,68,70,72,74,76,78,80,82,88,90,92,94,98,101,103,105,107],{"type":20,"html":21},"p","I run Claude Code most days — for writing, for refactors, for the kind of grunt edits that used to live in a Jupyter notebook. The bill adds up. Over the last week I started running a plugin called caveman mode, and watched my output token usage drop by about three-quarters while the actual work stayed identical.",{"type":20,"html":23},"This post is what caveman actually is, how to install it from GitHub, what it cuts vs. what it preserves, and the two places where I switch it back off.",{"type":25,"html":26},"h2","What "caveman mode" is",{"type":20,"html":28},"Caveman is a Claude Code plugin that rewrites the assistant's response style — not its reasoning. Articles (a\u002Fan\u002Fthe), filler (just\u002Freally\u002Fbasically\u002Factually), pleasantries (sure\u002Fcertainly\u002Fhappy to), and hedging all get dropped. Fragments are allowed. Synonyms get shorter ("fix" instead of "implement a solution for"). The technical substance — file paths, function names, error strings, code blocks — stays exact.",{"type":20,"html":30},"It registers a SessionStart hook, so it activates the moment you open a session and persists across turns. Three intensity levels ship by default: lite, full, and ultra. A wenyan variant compresses further by leaning on Classical Chinese grammar.",{"type":32,"emoji":33,"html":34},"callout","💡","The compression is in the output, not the input. Your prompt is not touched. The model still reasons normally — it just stops padding the reply.",{"type":25,"html":36},"Why the fluff is expensive",{"type":20,"html":38},"A typical Claude Code response wraps the answer in social glue: "Sure! I'd be happy to help. The issue you're experiencing is..." That intro is roughly fifteen tokens before any technical content appears. Multiply across hundreds of turns a week and you have a real line item on the bill.",{"type":20,"html":40},"The before\u002Fafter the README quotes is roughly:",{"type":42,"lang":43,"text":44},"code","plain text","\nNot: \"Sure! I'd be happy to help you with that. The issue\n      you're experiencing is likely caused by...\"\nYes: \"Bug in auth middleware. Token expiry check use `\u003C`\n      not `\u003C=`. Fix:\"",{"type":20,"html":46},"Same information. ~75% fewer tokens. The model thinks the same thoughts; you just stop paying for the wrapping paper.",{"type":25,"html":48},"Install",{"type":20,"html":50},"Caveman ships as a Claude Code plugin and lives on GitHub. Install from a session prompt:",{"type":42,"lang":52,"text":53},"bash","\u002Fplugin install caveman",{"type":20,"html":55},"Then any of these in-session:",{"type":42,"lang":52,"text":57},"\u002Fcaveman           # toggle on (full level)\n\u002Fcaveman lite      # lighter compression\n\u002Fcaveman ultra     # maximum compression\n\u002Fcaveman-stats     # show real token savings from session log\n\u002Fcaveman-commit    # write compressed conventional commit\n\u002Fcaveman-review    # one-line PR review comments",{"type":20,"html":59},"Typing "stop caveman" or "normal mode" reverts. Level persists until the session ends.",{"type":25,"html":61},"What gets dropped, what stays",{"type":63,"items":64},"ul",[65,66,67],"Dropped: articles, hedging, pleasantries, filler intensifiers, trailing "let me know if…" sign-offs.","Preserved: code blocks (untouched), error messages (quoted exact), file paths, function\u002Fclass names, command names, URLs.","Auto-disabled: security warnings, irreversible-action confirmations, multi-step sequences where fragment order risks misread.",{"type":20,"html":69},"That last bullet matters. Caveman quietly turns itself back on after the careful part is done. You do not have to babysit the toggle.",{"type":25,"html":71},"Real numbers from one week",{"type":20,"html":73},"I tracked usage via \u002Fcaveman-stats, which reads the session log directly — no model estimation, no marketing math. A representative day on a mid-sized refactor:",{"type":42,"lang":43,"text":75},"sessions:        14\noutput tokens:   ~118k   (estimated baseline ~430k)\nsavings:         ~73%\ncaveman level:   full",{"type":20,"html":77},"That lines up with the README's headline of ~75% within rounding. Input tokens are unchanged — your prompts, file reads, and tool results are not compressed. Output tokens are where the win lives, and output is the more expensive direction on most pricing tiers.",{"type":25,"html":79},"Where it bites in the GitHub flow",{"type":20,"html":81},"Three places I let caveman drive the work end-to-end:",{"type":83,"items":84},"ol",[85,86,87],"\u002Fcaveman-commit — generates Conventional Commit messages. Subject ≤50 chars, body only when the "why" is not obvious from the diff. No more "feat: add new functionality to handle the case where the user..."","\u002Fcaveman-review — PR review comments collapse to one line each: `path:line: severity: problem. fix.` Much easier to scan a 30-comment review.","\u002Fultrareview — multi-agent cloud review of the current branch. Not strictly caveman, but the output format is similarly compressed and pairs well.",{"type":20,"html":89},"Commits and PR descriptions land in normal prose — caveman is for conversation, not for artifacts another human reads out of context.",{"type":25,"html":91},"Where I switch it off",{"type":20,"html":93},"Two cases:",{"type":63,"items":95},[96,97],"Pair-debugging with someone watching my terminal. Fragment-style replies read as terse to another human in a way they do not read to me alone.","Documentation drafts and onboarding writeups. Anything that becomes Markdown someone else reads cold needs the articles put back.",{"type":99,"html":100},"blockquote","A communication style optimized for one reader is not free when there are two.",{"type":25,"html":102},"Verdict",{"type":20,"html":104},"Caveman pays for itself the first day. Output stays technically correct, the bill drops, the only adjustment is reading fragments instead of paragraphs in your own session. If you live in Claude Code, install it.",{"type":106},"hr",{"type":20,"html":108},"Next post: same measurement applied to \u002Floop and the cron-based scheduling routines — to see if scheduled background work has the same headroom.",[110,119,127,135],{"slug":111,"title":112,"date":113,"dateNice":114,"dateShort":115,"lastEdited":113,"excerpt":116,"tags":117,"readTime":118,"emoji":17,"cover":17},"installing-claude-code-with-one-command","Installing Claude Code with one command","2026-06-16","June 16, 2026","Jun 16","Setting up a coding tool usually means a checklist of installs before you even start. I wrote a one-command script that handles the whole thing — with a preview mode — for non-technical users.",[12,13],3,{"slug":120,"title":121,"date":122,"dateNice":123,"dateShort":124,"lastEdited":122,"excerpt":125,"tags":126,"readTime":118,"emoji":17,"cover":17},"deepseek-monitor-github-pull-requests","I let DeepSeek read every pull request so I don't have to","2026-06-04","June 4, 2026","Jun 4","AI made writing code cheap, but reviewing it still eats my mornings. So I built a small tool that hands the first-pass pull request review to DeepSeek for about a dollar a day.",[12,13,14],{"slug":128,"title":129,"date":9,"dateNice":130,"dateShort":131,"lastEdited":9,"excerpt":132,"tags":133,"readTime":134,"emoji":17,"cover":17},"deepseek-v4-pro-vs-claude-code","DeepSeek V4 Pro: 519 million tokens for $7","May 16, 2026","May 16","Our four-dev team blew through Claude Code's weekly budget, so we pointed everything at DeepSeek V4 Pro instead. Twelve days and $7 later, here's how it stacks up against Opus 4.7 and Sonnet 4.6 — and why the timing matters.",[12,13,14],4,{"slug":4,"title":5,"date":6,"dateNice":7,"dateShort":8,"lastEdited":9,"excerpt":10,"tags":136,"readTime":15,"emoji":16,"cover":17},[12,13,14],1781578540838]