RIP Reverse Engineer Anything!この 8 KB スキルが REA を上回る(38% 高速)
AICodeKing が、8 KB のスキルファイルである LeanREA を REA(Reverse Engineer Anything)とベンチマークし、REA の内部を説明し、36 回の実行で速度、コスト、精度を比較します。
この事例について
このチャンネルに参加して特典を利用する: https://www.youtube.com/channel/@aicodeking/join
LeanREA にスターを付ける:https://github.com/aicodeking/leanrea
このビデオでは、MCP サーバーや追加ツールなしで REA(Reverse Engineer Anything)と同じ仕事をするために私が構築した、単一の 8 KB スキルファイルである LeanREA を紹介します。REA の内部動作を説明し、次にスキルの内容を順に説明します:ルール、各ターゲットタイプを最速のコマンドにマッピングするルーターテーブル、すぐに実行できる最初の動き。次に、Claude Code で Sonnet 5.5 を使用した 36 回の実行にわたるベンチマークを示します。LeanREA は同じ精度で REA より 38% 高速で 35% 安価でした。さらにコンテキストテストとツールごとの速度テストも行いました。ベンチマークの限界、自分のスキルで見つけた 2 つのバグ、1 つの npx skills コマンドでインストールする方法についても話します。
セットアップ
インストール方法は元のソースをご確認ください。
原文 / README
RIP Reverse Engineer Anything! This 8 KB Skill Outperforms REA (38% Faster)
Join this channel to get access to perks: https://www.youtube.com/channel/@aicodeking/join
Star LeanREA here: https://github.com/aicodeking/leanrea
In this video, I'll show you LeanREA, a single 8 KB skill file that I built to do the same job as REA (Reverse Engineer Anything) without an MCP server or any extra tools. I explain what REA is under the hood, then go through what is inside the skill: the rules, the router table that maps each target type to its fastest command, and the ready-to-run first moves. Then I show my benchmark in Claude Code with Sonnet 5.5 across 36 runs, where LeanREA was 38% faster and 35% cheaper than REA at the same accuracy, plus the context test and a tool-by-tool speed test. I also talk about the limits of the benchmark, two bugs I found in my own skill, and how to install it with one npx skills command.
-- Key Takeaways:
📦 LeanREA is one 8 KB skill file with no MCP server and no extra tools; the agent runs the real tools in its normal shell. 🔀 REA is an MCP server plus a CLI that adds 207 tools, and under the hood it mostly calls tools you already have, like asar, otool, nm and jadx. 🧠 The skill has a short list of rules, a router table for 15 target types, and first moves that do many checks in one shell call. 📊 In 36 graded runs, LeanREA took 388 seconds, no skill took 479 seconds and REA took 627 seconds. 💸 LeanREA cost $2.17 against $3.36 for REA, with 56% fewer turns and 60% fewer tokens, and scored 74/75 against REA's 75/75. ⚠️ With tool search turned off, one "reply with OK" message with REA used about 400k tokens and cost $1.60. ⚡ On the same file, REA's Go build info took about half a second, and plain go version -m took 0.06 seconds. 👍 The plain model with no skill was already very good, so the big win is against REA, not against the model alone. 🔍 Two bugs I found in my own skill: ripgrep skips hidden folders like .webpack (use -uu), and zsh breaks echo with two equal signs. ✅ My verdict: if you are a programmer, use the small skill, because it is faster, cheaper, and you can read all of it in two minutes.
