juliusbrussee/caveman
9 skills777k installs630k starsGitHub
Install
npx skills add https://github.com/juliusbrussee/cavemanSkills in this repo
1Caveman CommitAgent-oriented skill that writes minimal, exact Conventional Commits—tight subjects and optional 72-column bodies focused on why, not what.140kinstalls2Caveman ReviewFormats code review as single actionable lines: location, problem, fix, with minimal noise.139kinstalls3Caveman CompressCompresses natural-language agent memory files for token savings via scripted caveman transformation.137kinstalls4Caveman HelpEphemeral Caveman command reference invoked via /caveman-help or natural phrases, output in caveman style only for that response.133kinstalls5CavecrewCavecrew is a delegation playbook for Julius Brussee’s caveman-style Claude Code subagents. Solo and indie builders use it when a long session is burning context on navigation, small fixes, or review loops that could run in a child agent instead of the main thread. The skill does not implement edits itself; it tells you when to launch cavecrew-investigator to find definitions and call sites, cavecrew-builder for obvious one- or two-file changes, or cavecrew-reviewer to scan a diff for bugs. It contrasts each preset with Anthropic defaults so you know when vanilla Explore or Code Reviewer is worth the fuller prose. Because subagents return caveman-compressed tool results, each delegation injects roughly sixty percent less text back into main context than uncompressed agents, which matters when you chain many lookups across an afternoon build. Reach for it when triggers like delegate to subagent, use cavecrew, save context, or compressed agent output match the task; skip spawning when you already know the one-line answer or the change cuts across many files.90.7kinstalls6Caveman StatsCaveman-stats is a hook-driven Claude Code skill that prints honest token accounting for the active session. Indie builders who compress output with caveman modes need receipts that do not come from the model’s own guesswork; this skill delegates math to hooks/caveman-stats.js via the caveman-mode-tracker, which intercepts the slash command and returns a blocked-decision payload the user sees immediately. Each run surfaces turn count, measured input and output tokens, an estimated non-caveman baseline, and derived savings—illustrated in the docs with forty-seven turns and on the order of sixty-five percent output reduction. Because ingestion is tied to the session log on disk, it fits anytime you are operating a long agent session and want to validate that compression is paying off before you ship another heavy delegation chain. It also persists lifetime savings for the statusline badge so micro-savings compound across days. Triggers are explicit: /caveman-stats. There is nothing to author in chat—the hook owns the pipeline end to end.90.6kinstalls7CompressCaveman compress is a journey-wide agent skill for solo builders who pay per token on long-running Claude Code or similar sessions. It targets natural-language memory files such as CLAUDE.md, todo lists, and preference docs—turning verbose prose into dense caveman-speak while keeping technical substance, code blocks, URLs, and structure intact. A local Python script pipeline detects file type, calls Claude to compress, validates output, and applies targeted fixes with up to two retries before failing safe on the original file. The compressed file replaces the original; a .original.md backup preserves the readable version. Use whenever context files grow heavy across Idea-through-Operate work, not as a substitute for deleting obsolete rules you no longer need.46.2kinstalls8Caveman Escaveman-es is a Spanish-language agent skill that switches replies into concise “caveman technical” prose so solo builders burn fewer tokens per turn without losing substance. You invoke it when you want less verbosity from your agent—modo cavernícola, habla como cavernícola, menos tokens, sé breve, or the /caveman-es command—with full as the default and lite or ultra for lighter or denser compression. The rules remove articles, filler, courtesy lines, and hedging while preserving exact technical terms, unchanged code blocks, and verbatim errors. Lite stays professional and complete sentences; full drops articles and allows fragments; ultra aggressively abbreviates domain shorthand and causal arrows. It is ideal for long debugging or review threads where clarity and token cost both matter, and it is not a substitute for structured planning skills when you still need full specs.516installs9Caveman CnCaveman CN is a journey-wide communication skill for solo builders who prefer Mandarin answers with minimal filler while keeping engineering precision. Invoke it when you say 中文 caveman, 穴居人模式, 更简短, 少点字, 用中文 caveman, call /caveman-cn, or explicitly want token-saving Chinese dialogue. The agent drops courtesy openers and vague transitions, favors short clauses or fragmented phrases, and follows a compact problem-action-reason-next-step shape. Lite mode trims fluff only; full mode is the default short-sentence style; ultra mode pushes phrasing toward arrow chains and word chunks. Git commits, PR descriptions, and code stay in normal form; only risky or irreversible guidance returns to full clarity. It helps across idea research, validation spikes, long build/debug threads, ship reviews, and operate incidents whenever chat verbosity costs time or context window. It is beginner-friendly to trigger but works best when you already read technical Chinese comfortably.441installs