![boboga777's tweet photo. [Open Source] I built a tool that lets you READ and REVIEW AI-generated code — even if you can't read code.
Vibe coding is everywhere now. Non-engineers (me included) are shipping real apps with AI. But there's a problem nobody solved:
"The AI wrote my entire app. I can't read a single line of it. Should I really ship this?"
That fear is justified. Independent testing found that roughly 45% of AI-generated code contains known security flaws (Veracode). We've all seen the incidents — production databases deleted, apps shipped with their user data wide open. The AI can build it. But there was no way for the person who asked for it to check it.
So I built one. It's called Code Interpreter (コード通訳), and it's open source as of today.
▼ What it does
It splits your code into small "units of meaning" (a few lines each) and shows a plain-English explanation next to the actual code — like a bilingual book. Then you review each unit with three buttons:
✓ Looks right (this is what I asked for)
✗ Wrong (I never asked for this)
? Not sure
It's not a quiz. You're just answering: "is this what I wanted?"
▼ Every feature
- Press a file name → that file appears, already translated (translation is pre-generated, so it's instant)
- Risky spots — money, personal data, deletion — get a ⚠ Caution mark with an amber explanation of what could go wrong. A floating button tours only the ⚠ spots
- Hover any line of code for its one-line translation; hover keywords like `if` and `const` for hints
- Write free comments in ordinary language ("add a confirmation screen before sending")
- Your ✗ answers and comments are saved to a file. Tell the AI "fix the flagged items" in chat and it fixes them in bulk
- A verdict at the bottom: it turns green ONLY when every unit is marked ✓, every file was read, and nothing critical was detected
- When the code changes, all answers auto-expire — approvals of old code are never recycled for new code
- Learning mode: explanations are hidden, you guess what the code does, then click to check. Flashcard-style active recall — you actually learn to read code by reviewing your own app
- Works in English (--lang en) and Japanese
▼ The design principle I care about most
This tool never says "your code is safe."
Unread files are shown in red. AI explanations can be wrong, so the real code always sits right next to them. The verdict won't turn green while anything is unanswered or unverified. What it gives you is not reassurance — it's an honest record of what YOU reviewed and what remains unchecked.
Polished reports make people careless. So I built the opposite.
▼ Getting started (3 steps)
1. git clone (see README)
2. code-translate your-app --open --lang en
3. Your browser opens. From then on, just double-click the launcher file it drops in your folder
Works standalone or with Claude Code (~$0.1–0.3 per translation for a small app; unchanged code is never re-translated).
If you're vibe coding anything real — or reviewing what an AI or contractor shipped you — I'd love your feedback. Stars appreciated more than you know.
GitHub: https://t.co/STaK9oNqSj
English manual (with screenshots): https://t.co/AIqR1QZ2Ne
日本語マニュアル: https://t.co/ffuctJXSuX
#OpenSource #VibeCoding #AIcoding #BuildInPublic #ClaudeCode](https://pbs.twimg.com/media/HQj0APIbEAAihGp.png)
![boboga777's tweet photo. [Open Source] I built a tool that lets you READ and REVIEW AI-generated code — even if you can't read code.
Vibe coding is everywhere now. Non-engineers (me included) are shipping real apps with AI. But there's a problem nobody solved:
"The AI wrote my entire app. I can't read a single line of it. Should I really ship this?"
That fear is justified. Independent testing found that roughly 45% of AI-generated code contains known security flaws (Veracode). We've all seen the incidents — production databases deleted, apps shipped with their user data wide open. The AI can build it. But there was no way for the person who asked for it to check it.
So I built one. It's called Code Interpreter (コード通訳), and it's open source as of today.
▼ What it does
It splits your code into small "units of meaning" (a few lines each) and shows a plain-English explanation next to the actual code — like a bilingual book. Then you review each unit with three buttons:
✓ Looks right (this is what I asked for)
✗ Wrong (I never asked for this)
? Not sure
It's not a quiz. You're just answering: "is this what I wanted?"
▼ Every feature
- Press a file name → that file appears, already translated (translation is pre-generated, so it's instant)
- Risky spots — money, personal data, deletion — get a ⚠ Caution mark with an amber explanation of what could go wrong. A floating button tours only the ⚠ spots
- Hover any line of code for its one-line translation; hover keywords like `if` and `const` for hints
- Write free comments in ordinary language ("add a confirmation screen before sending")
- Your ✗ answers and comments are saved to a file. Tell the AI "fix the flagged items" in chat and it fixes them in bulk
- A verdict at the bottom: it turns green ONLY when every unit is marked ✓, every file was read, and nothing critical was detected
- When the code changes, all answers auto-expire — approvals of old code are never recycled for new code
- Learning mode: explanations are hidden, you guess what the code does, then click to check. Flashcard-style active recall — you actually learn to read code by reviewing your own app
- Works in English (--lang en) and Japanese
▼ The design principle I care about most
This tool never says "your code is safe."
Unread files are shown in red. AI explanations can be wrong, so the real code always sits right next to them. The verdict won't turn green while anything is unanswered or unverified. What it gives you is not reassurance — it's an honest record of what YOU reviewed and what remains unchecked.
Polished reports make people careless. So I built the opposite.
▼ Getting started (3 steps)
1. git clone (see README)
2. code-translate your-app --open --lang en
3. Your browser opens. From then on, just double-click the launcher file it drops in your folder
Works standalone or with Claude Code (~$0.1–0.3 per translation for a small app; unchanged code is never re-translated).
If you're vibe coding anything real — or reviewing what an AI or contractor shipped you — I'd love your feedback. Stars appreciated more than you know.
GitHub: https://t.co/STaK9oNqSj
English manual (with screenshots): https://t.co/AIqR1QZ2Ne
日本語マニュアル: https://t.co/ffuctJXSuX
#OpenSource #VibeCoding #AIcoding #BuildInPublic #ClaudeCode](https://pbs.twimg.com/media/HQjz8uSboAAbXz7.jpg)
![boboga777's tweet photo. [Open Source] I built a tool that lets you READ and REVIEW AI-generated code — even if you can't read code.
Vibe coding is everywhere now. Non-engineers (me included) are shipping real apps with AI. But there's a problem nobody solved:
"The AI wrote my entire app. I can't read a single line of it. Should I really ship this?"
That fear is justified. Independent testing found that roughly 45% of AI-generated code contains known security flaws (Veracode). We've all seen the incidents — production databases deleted, apps shipped with their user data wide open. The AI can build it. But there was no way for the person who asked for it to check it.
So I built one. It's called Code Interpreter (コード通訳), and it's open source as of today.
▼ What it does
It splits your code into small "units of meaning" (a few lines each) and shows a plain-English explanation next to the actual code — like a bilingual book. Then you review each unit with three buttons:
✓ Looks right (this is what I asked for)
✗ Wrong (I never asked for this)
? Not sure
It's not a quiz. You're just answering: "is this what I wanted?"
▼ Every feature
- Press a file name → that file appears, already translated (translation is pre-generated, so it's instant)
- Risky spots — money, personal data, deletion — get a ⚠ Caution mark with an amber explanation of what could go wrong. A floating button tours only the ⚠ spots
- Hover any line of code for its one-line translation; hover keywords like `if` and `const` for hints
- Write free comments in ordinary language ("add a confirmation screen before sending")
- Your ✗ answers and comments are saved to a file. Tell the AI "fix the flagged items" in chat and it fixes them in bulk
- A verdict at the bottom: it turns green ONLY when every unit is marked ✓, every file was read, and nothing critical was detected
- When the code changes, all answers auto-expire — approvals of old code are never recycled for new code
- Learning mode: explanations are hidden, you guess what the code does, then click to check. Flashcard-style active recall — you actually learn to read code by reviewing your own app
- Works in English (--lang en) and Japanese
▼ The design principle I care about most
This tool never says "your code is safe."
Unread files are shown in red. AI explanations can be wrong, so the real code always sits right next to them. The verdict won't turn green while anything is unanswered or unverified. What it gives you is not reassurance — it's an honest record of what YOU reviewed and what remains unchecked.
Polished reports make people careless. So I built the opposite.
▼ Getting started (3 steps)
1. git clone (see README)
2. code-translate your-app --open --lang en
3. Your browser opens. From then on, just double-click the launcher file it drops in your folder
Works standalone or with Claude Code (~$0.1–0.3 per translation for a small app; unchanged code is never re-translated).
If you're vibe coding anything real — or reviewing what an AI or contractor shipped you — I'd love your feedback. Stars appreciated more than you know.
GitHub: https://t.co/STaK9oNqSj
English manual (with screenshots): https://t.co/AIqR1QZ2Ne
日本語マニュアル: https://t.co/ffuctJXSuX
#OpenSource #VibeCoding #AIcoding #BuildInPublic #ClaudeCode](https://pbs.twimg.com/media/HQjz54qbIAE7wll.jpg)
Olivia
Online
沖縄AI実践ラボ
@boboga777
🌺沖縄AI実践ラボ｜思いついたらAIへ。AIエージェントを活用した業務自動化・システム構築を実践。iOSアプリを多数開発し、作り方も指導します。GitHubで業務フロー自動化OSSを公開中。Discord・Zoom交流会・オフ会を開催予定。AI活用の相談歓迎📩
Joined November 2021
54 Following
65 Followers
474 Posts
これガチでマジだから。
Spot on — those three are exactly what the ⚠ risky-spots pass targets. Two notes from building it:
① "Auth in comments, not code" is the one people miss most — the doc says it's handled, so the brain autocompletes "done." We treat that as a named blind spot: a finding can't go green until you open the actual code path, not the comment.
② "Destructive ops without confirmation" literally bit the audit tool itself — the 3-AI review caught our own bundle step deleting a directory with no guard. The pattern really is everywhere.
The whole design is built so a non-engineer gives an explicit allow/deny on exactly those spots, and it won't show green while any is unresolved. Would genuinely love your eyes on the critical-review mode.
[Open Source] I built a tool that lets you READ and REVIEW AI-generated code — even if you can't read code.
Vibe coding is everywhere now. Non-engineers (me included) are shipping real apps with AI. But there's a problem nobody solved:
"The AI wrote my entire app. I can't read a single line of it. Should I really ship this?"
That fear is justified. Independent testing found that roughly 45% of AI-generated code contains known security flaws (Veracode). We've all seen the incidents — production databases deleted, apps shipped with their user data wide open. The AI can build it. But there was no way for the person who asked for it to check it.
So I built one. It's called Code Interpreter (コード通訳), and it's open source as of today.
▼ What it does
It splits your code into small "units of meaning" (a few lines each) and shows a plain-English explanation next to the actual code — like a bilingual book. Then you review each unit with three buttons:
✓ Looks right (this is what I asked for)
✗ Wrong (I never asked for this)
? Not sure
It's not a quiz. You're just answering: "is this what I wanted?"
▼ Every feature
- Press a file name → that file appears, already translated (translation is pre-generated, so it's instant)
- Risky spots — money, personal data, deletion — get a ⚠ Caution mark with an amber explanation of what could go wrong. A floating button tours only the ⚠ spots
- Hover any line of code for its one-line translation; hover keywords like `if` and `const` for hints
- Write free comments in ordinary language ("add a confirmation screen before sending")
- Your ✗ answers and comments are saved to a file. Tell the AI "fix the flagged items" in chat and it fixes them in bulk
- A verdict at the bottom: it turns green ONLY when every unit is marked ✓, every file was read, and nothing critical was detected
- When the code changes, all answers auto-expire — approvals of old code are never recycled for new code
- Learning mode: explanations are hidden, you guess what the code does, then click to check. Flashcard-style active recall — you actually learn to read code by reviewing your own app
- Works in English (--lang en) and Japanese
▼ The design principle I care about most
This tool never says "your code is safe."
Unread files are shown in red. AI explanations can be wrong, so the real code always sits right next to them. The verdict won't turn green while anything is unanswered or unverified. What it gives you is not reassurance — it's an honest record of what YOU reviewed and what remains unchecked.
Polished reports make people careless. So I built the opposite.
▼ Getting started (3 steps)
1. git clone (see README)
2. code-translate your-app --open --lang en
3. Your browser opens. From then on, just double-click the launcher file it drops in your folder
Works standalone or with Claude Code (~$0.1–0.3 per translation for a small app; unchanged code is never re-translated).
If you're vibe coding anything real — or reviewing what an AI or contractor shipped you — I'd love your feedback. Stars appreciated more than you know.
GitHub: https://t.co/STaK9oNqSj
English manual (with screenshots): https://t.co/AIqR1QZ2Ne
日本語マニュアル: https://t.co/ffuctJXSuX
#OpenSource #VibeCoding #AIcoding #BuildInPublic #ClaudeCode
![boboga777's tweet photo. [Open Source] I built a tool that lets you READ and REVIEW AI-generated code — even if you can't read code.
Vibe coding is everywhere now. Non-engineers (me included) are shipping real apps with AI. But there's a problem nobody solved:
"The AI wrote my entire app. I can't read a single line of it. Should I really ship this?"
That fear is justified. Independent testing found that roughly 45% of AI-generated code contains known security flaws (Veracode). We've all seen the incidents — production databases deleted, apps shipped with their user data wide open. The AI can build it. But there was no way for the person who asked for it to check it.
So I built one. It's called Code Interpreter (コード通訳), and it's open source as of today.
▼ What it does
It splits your code into small "units of meaning" (a few lines each) and shows a plain-English explanation next to the actual code — like a bilingual book. Then you review each unit with three buttons:
✓ Looks right (this is what I asked for)
✗ Wrong (I never asked for this)
? Not sure
It's not a quiz. You're just answering: "is this what I wanted?"
▼ Every feature
- Press a file name → that file appears, already translated (translation is pre-generated, so it's instant)
- Risky spots — money, personal data, deletion — get a ⚠ Caution mark with an amber explanation of what could go wrong. A floating button tours only the ⚠ spots
- Hover any line of code for its one-line translation; hover keywords like `if` and `const` for hints
- Write free comments in ordinary language ("add a confirmation screen before sending")
- Your ✗ answers and comments are saved to a file. Tell the AI "fix the flagged items" in chat and it fixes them in bulk
- A verdict at the bottom: it turns green ONLY when every unit is marked ✓, every file was read, and nothing critical was detected
- When the code changes, all answers auto-expire — approvals of old code are never recycled for new code
- Learning mode: explanations are hidden, you guess what the code does, then click to check. Flashcard-style active recall — you actually learn to read code by reviewing your own app
- Works in English (--lang en) and Japanese
▼ The design principle I care about most
This tool never says "your code is safe."
Unread files are shown in red. AI explanations can be wrong, so the real code always sits right next to them. The verdict won't turn green while anything is unanswered or unverified. What it gives you is not reassurance — it's an honest record of what YOU reviewed and what remains unchecked.
Polished reports make people careless. So I built the opposite.
▼ Getting started (3 steps)
1. git clone (see README)
2. code-translate your-app --open --lang en
3. Your browser opens. From then on, just double-click the launcher file it drops in your folder
Works standalone or with Claude Code (~$0.1–0.3 per translation for a small app; unchanged code is never re-translated).
If you're vibe coding anything real — or reviewing what an AI or contractor shipped you — I'd love your feedback. Stars appreciated more than you know.
GitHub: https://t.co/STaK9oNqSj
English manual (with screenshots): https://t.co/AIqR1QZ2Ne
日本語マニュアル: https://t.co/ffuctJXSuX
#OpenSource #VibeCoding #AIcoding #BuildInPublic #ClaudeCode](https://pbs.twimg.com/media/HQj0CWSaIAALDiz.png)
【OSS公開】コードが読めなくても、AIの書いたコードを「読める・確認できる」ツールを作りました
AIにアプリを作らせる人が一気に増えました。私もその一人です。でも、ずっと引っかかっていたことがあります。
「AIが書いたコード、自分では1行も読めない。このまま公開していいの?」
これは気のせいではなくて、実測データがあります。AIが生成したコードの約45%に既知のセキュリティ問題が含まれる、という調査結果(Veracode)。動くものは作れる。でも「安全かどうかを発注者が確認する手段」が、世の中にほぼ存在しない。
だから作りました。「コード通訳」。今日からオープンソースです。
▼ 何をするツールか
コードを「数行の意味のまとまり」ごとに区切って、左に日本語の説明、右に実際のコードを並べます。英語の本の対訳版と同じ形式です。あなたがやることは、説明を読んで3つのボタンのどれかを押すだけ。
✓ 問題ない(意図どおり)
✗ おかしい(そんな動き頼んでない)
? わからない
正解を当てるテストではありません。「あなたが望んだ通りか」を答えるだけです。
▼ 機能ぜんぶ紹介
・ファイル名のボタンを押すと、そのコードが翻訳済みで即表示(翻訳は事前に済んでいるので待ち時間ゼロ)
・お金・個人情報・削除に触れる危ない箇所には⚠マーク。「次の注意箇所へ」ボタンで⚠だけを順番に巡回できる
・コードの行にマウスを乗せると、その1行の訳が出る。ifやconstなどの単語に触れると意味のヒントも
・気になることは「コメントを書く」に普通の言葉で。「ここは確認画面を挟んでほしい」とか
・✗とコメントはファイルに保存され、AI(Claude Code)に「たまった指摘を直して」と言うだけでまとめて修正される
・一番下に「公開してよいか」の判定。全部✓になるまで絶対に緑にならない
・コードが変わると回答は自動で無効化。「昔のコードへの承認」が使い回される事故を防ぐ
・学習モード: 説明を隠して「この数行は何をしている?」→自分で予想→クリックで答え合わせ。単語カード方式で、使うほどコードが読めるようになる
・AIが開発中なら、編集のたびに差分の対訳が自動で付いてくる「翻訳モード」も
・Gitの履歴を日本語のタイムラインにする「作業履歴の翻訳」で、後からの振り返りもできる
▼ 設計で一番こだわったこと
このツールは「安全です」とは絶対に言いません。
読めていないファイルは「未読」と赤で表示する。AIの説明は間違いうるから、必ず実物のコードを隣に置く。未回答・✗・?が残る限り判定は緑にしない。売っているのは「安心」ではなく、「あなたが何を確認して、何が未確認か」の正直な記録です。
きれいなレポートほど人を油断させる。だから「見せかけの安心」を作らない設計にしました。
▼ 使い方は3ステップ
1. git clone(READMEに手順あり)
2. code-translate 自分のアプリ --open
3. ブラウザが勝手に開く。次からはフォルダに自動生成される「コード通訳を開く.command」をダブルクリックするだけ
Claude Codeユーザーなら、チャットに「コードを翻訳して」と打つだけでも動きます。費用は翻訳の生成時のみ(小さいアプリで1回数十円)。コードに変更がなければ何度でも無料。
▼ こんな人に
・AIでアプリを作っているけど、中身がブラックボックスで不安な人
・外注やAIの納品物を「発注者として」確認したい人
・コードを読めるようになりたいけど、教材より「自分のアプリ」で学びたい人
リポジトリはこちら。スターをもらえたら泣いて喜びます。フィードバックはIssueへどうぞ。
GitHub: https://t.co/STaK9oNqSj
日本語マニュアル(スクショ付き5分): https://t.co/ffuctJXSuX
English manual: https://t.co/AIqR1QZ2Ne
#OSS #バイブコーディング #ClaudeCode #個人開発 #AIコーディング
More and more people are having AI build their apps. But—
**"It runs. Yet is it doing anything weird inside? Is it actually safe? I can't tell for myself."**
I've open-sourced a tool that answers that worry head-on. It's called **ai-audit**.
The core idea is simple:
👉 **Don't let the AI that wrote the code check its own work.**
👉 **Bring in "eyes other than the maker's" — a different company's AI (OpenAI's Codex, xAI's Grok) audits it with a stranger's eyes.**
Even humans miss the typos in their own writing while a stranger spots them instantly. AI is the same: the AI that wrote the code carries the assumption "I'm sure I got it right," and overlooks its own mistakes. So you show it to an AI that *isn't* the author. We call this **cognitive independence.**
—
And here's the proof I most want to share.
**Before launching, I had three AIs audit ai-audit itself.**
They found defects the implementing AI (me, Claude) had missed — each with reproduction steps. And the striking part:
**Grok and Codex found completely *different* holes.**
🟣 What Grok (xAI) found
· A deletion-permission check that could be spoofed with a single marker file
· A "highest severity" vuln (CVSS 9.8) misread as "low" and missed
🟢 What Codex (OpenAI) found
· A bypass when the input path is a shortcut (symlink)
· Broken/forged scan results read as "all clear"
· A scan labelled "whole history" that actually only looked at current files
· An empty "seal" that could be created covering zero files
Meanwhile, in the very first round, **both AIs agreed on the same critical bugs** (like a data-loss risk).
So:
**Everyone flags the big ones; they divide up the subtle ones.**
The more different-vendor AIs audit it, the more coverage adds up. That's the clearest evidence for "several AIs beat one." With a single AI, half of those holes would still be there.
I fixed them all, added 102 regression tests and machine gates, and left the entire trail in the repo. **The fact that my own audit tool keeps getting holes pointed out by independent AIs — and that I show all of it — is the most honest evidence that independent auditing actually works.**
—
**And the other half of it: don't stop at "found."**
Detection alone doesn't make code better. The heart of ai-audit is a **convergence loop**: **fix → a different AI re-audits → fix again**, repeated **until no serious findings remain.**
I demonstrated this too. It didn't finish in one pass before launch:
**Round 1** — 4 critical bugs → fixed → **Round 2** (a different AI, Grok) found residual holes *inside my round-1 fixes* → fixed → **Round 3** (Codex) found finer ones → fixed. **Each round, the independent AIs found fewer — it converged.**
The catch: **"loop until the checker goes quiet" over-fits to that AI's blind spots and backfires.** So there are four rules:
① **Separate who fixes from who judges** (Claude implements; a different-vendor AI judges)
② **Rotate the auditor each round** (don't stop at the first green; stop when independent passes come back clean in a row)
③ **Record *why* each finding was found**
④ **Stop honestly** (never weaken a test or gate just to pass)
**Detection (several independent AIs) × convergence (loop until it passes, correctly).** That product is the difference between a one-shot review that just points at things and a system that actually raises quality.
—
**Who is it for?**
"An AI built it" earns no trust from executives or clients. What you need is a deliverable that explains, with evidence, *what was checked against which standard, how far, and what remains unverified.* ai-audit generates exactly that — a technical audit report linking requirement IDs, tests, and evidence. It turns a verbal "I checked it" into a traceable document at delivery time. In any setting where your work is audited or inspected, having that explainable trail is what earns trust.
And **reviewing code isn't easy for non-engineers.** This tool takes over that "checking" work — not by trusting one AI, but by using several independent, different-vendor AIs.
**Two AIs is fine; three is recommended.**
· Claude + Codex (two) = officially supported minimum
· + Cursor (Grok) (three) = recommended; rotate auditors and catch more *different* holes
—
⚠️ Honestly: this is not magic that guarantees "safe." Its job is to show, without hiding anything, what was checked and what is still unverified — and never to relabel "not tested" as "safe." It's currently a **v1.8 Preview with limitations stated openly.**
—
📘 The gentlest guide (for non-engineers)
https://t.co/zr3QWToSEp
💻 GitHub (MIT, free). If it helps, a ⭐Star is hugely appreciated
https://t.co/ixMEhsNr69
Three steps: ①have your AI read the repo →②build your app as usual →③say "audit the code."
#AI #GenAI #Claude #ChatGPT #Codex #Grok #Cybersecurity #OpenSource #IndieHackers #vibecoding #CodeReview #AIagents #AppSec
AIにアプリを作らせる人が急増しています。でも——
**「動いてはいる。けど、中で変なことをしていないか、本当に大丈夫なのか、自分では分からない」**
このモヤモヤに正面から答えるオープンソースを公開しました。
その名は **ai-audit**。
コアの発想はシンプルです。
👉 **コードを書いたAI自身にチェックさせない。**
👉 **"作った本人以外の目"——別の会社の別のAI(OpenAIのCodex、xAIのGrok)に、他人の目で監査させる。**
人間でも、自分の書いた文章の誤字は見つけにくく、他人はすぐ見つけますよね。AIも同じで、作った本人は「ちゃんとできたはず」という思い込みのまま自分のミスを見逃します。だから"当事者じゃないAI"に見せる。これを認知的独立性と呼んでいます。
—
そして今回、いちばん伝えたい実証があります。
**公開前に、この ai-audit 自身を3つのAIに監査させました。**
実装した私(Claude)が見落としていた欠陥を、他社のAIたちが次々と、しかも再現手順つきで見つけました。しかも——
**GrokとCodexは、まったく"別々の穴"を見つけたのです。**
🟣 Grok(xAI)が見つけた穴
・削除の許可判定を、目印ファイル1枚で偽装できる欠陥
・危険度「最高(CVSS 9.8)」の脆弱性を"低"と誤判定して見逃す欠陥
🟢 Codex(OpenAI)が見つけた穴
・入力がショートカット(symlink)のときのすり抜け
・壊れた/偽装された検査結果を"問題なし"と誤読
・「全履歴を検査」と表示しつつ、実は今のファイルしか見ていなかった
・中身ゼロの"空っぽの封印"が作れてしまう
一方で、最初の監査では両者が**同じ重大バグ(データ消失の危険など)には揃って警告**しました。
つまり——
**重要なものは全員一致で、細かいものは手分けして発見。**
複数の別会社AIに監査させるほど、カバレッジは"足し算"で広がる。これが「1つのAIより複数」の何よりの証拠です。1つしか使っていなかったら、上の穴の半分は残ったままでした。
これらの欠陥は全て修正し、再発防止の回帰テスト(102件)と機械ゲートを追加。その過程も全部リポジトリに残してあります。**「監査ツール自身が、別系統AIの監査でまだ穴を指摘される」——それを隠さず全部見せることが、独立監査に実利がある最も正直な証拠だと考えています。**
—
**そして、もう一つの肝は「見つけて終わり」にしないこと。**
検出だけでは、コードは良くなりません。ai-audit の本体は、**直す → 別のAIがまた監査する → また直す**を、**重大な指摘が出なくなるまで繰り返す"収束ループ"**です。
これも実演しました。今回の公開前も、1回では終わりませんでした。
**1回目**の監査で重大バグ4件 → 修正 → **2回目**(別AI=Grok)で"1回目の修正の中の残り穴" → 修正 → **3回目**(Codex)でさらに細かい残り → 修正。**回すたびに、独立したAIが見つける穴は減っていき、収束していきました。**
ポイントは、**「チェックする側が黙るまで回す」と、そのAIの盲点に過剰適合して逆効果**になること。だから4つの掟を守ります:
①**直す担当と、判断する担当を分ける**(実装はClaude、判断は別会社AI)
②**監査役を毎回交代**する(最初の緑で止めず、独立した複数回が連続クリーンで止める)
③**なぜ発見できたかを記録**する
④**正直に止める**(テストやゲートを弱めて無理やり通さない)
**検出(複数の別系統AI)× 収束(通るまで、正しく回す)** の掛け算。これが、ただ指摘するだけの単発レビューと、実際に品質を押し上げる仕組みの違いです。
—
**誰のための道具か。**
「AIで作りました」だけでは、経営層や取引先の信頼は得られません。必要なのは、**何を基準に・どこまで検証し・何が未検証かを、証拠つきで説明できる成果物**です。ai-audit はそれを、要求IDとテストと証拠を結んだ技術監査報告書として生成します。納品・検収の場で、口頭の「確認しました」を、追跡可能な書類に変えます。監査や審査を受ける現場では、この"説明できる証跡"の有無が信頼を分けます。
そして——**非エンジニアがコードレビューをするのは簡単ではない。** その"確かめる"作業を、一人のAIに丸投げせず、複数の別系統AIで肩代わりする。それがこの道具のねらいです。
**AIは2つでもOK、3つが推奨。**
・Claude + Codex(2つ)= 正式サポートの最小構成
・+ Cursor(Grok)(3つ)= 推奨。監査役を交代でき、上のように"違う穴"がより多く見つかる
—
⚠️ 正直に言います。これは「絶対安全」を保証する魔法ではありません。役目は「何を調べ、何がまだ未検証か」を隠さず示すこと。「未検証」を「安全」と言い換えない——この誠実さこそが価値です。現時点は**限界を明示したv1.8 Preview**です。
—
📘 いちばんやさしい使い方(非エンジニア向け・日本語)
https://t.co/5xVQtbGhsO
📕 詳しいマニュアル(全体)
https://t.co/QQ9isSU1vJ
💻 GitHub(MIT・無料)。役に立ったら ⭐Star で応援してください
https://t.co/ixMEhsNr69
使い方は3ステップだけ。①AIにリポジトリを読ませる →②いつも通りアプリを作る →③「コードを監査して」と言うだけ。
#AI #生成AI #Claude #ChatGPT #Codex #セキュリティ #オープンソース #個人開発 #vibecoding #コードレビュー #AIエージェント
Cloudflareについて、全23章・約23万字のHTML完全版をnoteにまとめました！
この投稿を8月23日（日）までにリポストすると、無料で全文読めます。
【約23万字】Cloudflare完全ガイド｜Web・セキュリティ・AI・MCPをこれ1本で理解するhttps://t.co/SKhmupf3Rt
https://t.co/mKfRMFhVFb
Cursor: Cloud Agents and Cursor Harness Improvements
What changed:
- Cloud agents can now subscribe to PRs, Slack threads, and scheduled tasks, waking automatically when something happens
- Subagents can run on their own virtual machines, each with an isolated copy of the project and clean context
- Use /goal to give an agent a long-lived objective to work toward until fully complete
- Steer the agent mid-run without interruption — follow-ups wait for the next tool call instead of cutting it off
Details in thread ↓
https://t.co/op4hdHNG6u
I’ve released an open-source Work Automation Atlas for freelancers, solo developers, internal automation teams, and small agencies.
The challenge in business automation is rarely just “Can AI generate this document?”
The real questions are:
• What work is actually being performed?
• Which parts can be standardized?
• What information and historical materials are required?
• Which exceptions must be handled?
• Who reviews and approves the result?
• What should never be fully automated?
• Is the workflow ready for a proof of concept?
The toolkit classifies client work using 25 reusable automation patterns evaluated against 122 government, enterprise, and manufacturing cases.
For example, if a client wants to automate a monthly report, the completed PDF alone is not enough.
You may also need:
• source data from the same reporting period
• matching historical reports
• calculation formulas
• exception-handling rules
• terminology and formatting requirements
• the person responsible for approval
• the final delivery channel
• rules for correcting or reissuing the report
Once those materials are identified, an AI agent can provide a much more useful answer during the first discovery meeting:
“This part can be tested now.”
“These materials are still missing.”
“This decision must remain with a human.”
“This is the realistic scope of the first proof of concept.”
Classification does not automatically mean that a workflow is easy to automate. It gives developers and clients a shared structure for discussing feasibility, missing evidence, risks, human approvals, and implementation scope before making promises.
The project can be used with Codex, Claude Code, and Cursor, helping AI agents analyze a workflow and prepare an initial automation diagnosis.
GitHub:
https://t.co/75Us3gKg9j
English manual:
https://t.co/NZAuXQff9c
#AI #Automation #WorkflowAutomation #BusinessAutomation #AIAgents #OpenSource #FreelanceDeveloper #DigitalTransformation
【業務自動化の相談で、最初に困ること】
お客様から
「この仕事、AIで自動化できますか？」
と聞かれても、何を確認すればよいか分からない。
そこで、AI Automation Starter Kitに新しい業務診断機能を追加しました。
業務内容を入力すると、
・25の基礎型から仕事を分類
・必要な過去資料や元データを提示
・AIと人の担当範囲を分離
・既存の自動化フローを推薦
・ローカルで試せるPoC案を作成
という流れで検討を進められます。
例えば月報業務なら、
「過去の完成月報」
「集計元のExcel」
「計算ルール」
「修正履歴」
「提出先と承認者」
があれば、どこまで自動化できるかを整理できます。
これがあると、お客様との面談で、
「この資料がそろえば、集計と下書きは自動化できます。最終確認と提出は人に残します」
と、具体的に説明できるようになります。
Codex、Claude Code、Cursorから読み込める共通スキルにもしています。
AI副業として企業へ業務自動化を提案したい人、個人開発者、社内の業務改善を任された人向けです。
専門用語を減らした操作マニュアルも公開しました。
▼やさしい操作マニュアル
https://t.co/m1ni7jTEPu
▼GitHub
https://t.co/75Us3gKg9j
#AI副業 #業務自動化 #生成AI #個人開発
Origin, our code hosting platform, is now live.
It's fast, easy to use, and deeply integrated with Cursor.
Get started by syncing your repos from GitHub.
ほとんどの人が、
Claude Codeの使い方を
勘違いしています。
ターミナルを開く。
↓
プロンプトを書く。
↓
あとはAIに丸投げ。
これだけでは、
Claude Codeの本領を
ほぼ引き出せません👇
本当に重要なのは、
プロンプトではなく
「仕組み」です。
Claude Codeは、
大きく4つの層で考えると
一気に化けます。
① CLAUDE.md
→ プロジェクトの記憶
・アーキテクチャ
・ルール
・コマンド
・開発上の慣習
毎回説明する情報を
ここに持たせる。
② Skills
→ 再利用できる知識・手順
・テスト手順
・コードレビュー
・デプロイ
・定型ワークフロー
毎回プロンプトで
長々と指示する必要が
なくなります。
③ Hooks
→ 確実に実行させる仕組み
・セキュリティチェック
・ルールの強制
・処理の自動化
AIの判断任せにせず、
決めた処理を動かせる。
④ Agents
→ 専門家を分業させる
複雑な仕事を分解して、
専門のサブエージェントに
役割を持たせる。
この4つを組み合わせると👇
CLAUDE.md
↓
Skills
↓
Hooks
↓
Agents
Claude Codeは
ただの「賢いチャット」から、
自分専用の
AI開発システムに変わります。
多くの人は、
いきなりプロンプトから
始めてしまう。
でも、
出力の質を変えるのは
プロンプトの上手さだけじゃない。
その前にある
「環境設計」です。
2026年、
AIエージェントを
本気で使うなら、
プロンプトを学ぶだけでは
足りなくなってきています。
これから重要になるのは、
「何を指示するか」より
「どう動く仕組みを作るか」。
これからのAI開発は、
プロンプト優先から
「システム優先」へ
画像を保存して
Claude codeを使いこなしてください👇
【速報】
Codex GPT-5.6 Solで
スクロールで物語が動くサイトを丸ごと作る手順が公開された😳
このワークフローの中身をまとめる👇
https://t.co/osLV1IdO8h
・映像素材はSeedanceとLTX-2.5で生成する
・その素材をCodex GPT-5.6 Solがサイト側へ組み込む
・スクロール量に合わせて映像とテキストが連動する
Googleが「未来予測AI」を無料配布してた...
いまめっちゃ話題になってるプロジェクトなんですが
Google Researchが開発した「TimesFM」という時系列予測AIで、1000億のタイムポイントで事前学習済み。
何がヤバいかというと、自分のデータを渡すだけで、ファインチューニングなしにゼロショットで売上・アクセス数・需要予測が出てくるんです。
使い方は拍子抜けするほど簡単。
インストールして、過去の数値データを渡して「予測して」と実行する。
たったこれだけです。
これが何をしてくれるかというと、過去のデータから「次にどうなるか」を確率分布つきで予測してくれます。
しかもICML 2024に採択された本物の論文付きで、パラメータたった200Mなのにローカルマシンでサクサク動く。
・1000億タイムポイントで学習済み → 未知のデータにそのまま適用可能
・200Mパラメータの軽量設計 → GPU無しのCPUでも動作
・BigQuery・Googleスプレッドシートにも統合済み → SQL一行で予測
・Apache-2.0ライセンス → 商用利用も完全無料
今超話題の未来予想AIはこちら👇
https://t.co/JExqCMa4Yx
SaaSサービスの紹介アニメーション動画
全てAIで制作
企画：ChatGPT
動画：Gemini Omni Flash
音楽：Gemini Omni Flash
音声：Gemini Omni Flash
編集：Gemini Omni Flash
https://t.co/DYhOigWiBq
MiniMax H3でモーショングラフィックをA100で回した現状版。タイポグラフィあり4本。H.264/AAC、640x352、24fps、243フレーム、10.125秒。
雰囲気はいい。全フレーム見ると、01 editorial cards、03 amber product、04 paper collage系の一部で、カード／ボトル／紙面がタイル状に反復して崩れてる。画質差は640x352だけじゃなくて、生成フレーム側の破綻もある。再生成も品質改善もまだしてない。
ちなみにこれ、GoogleのFLOWで制作してます！
無料アカウントでも毎日50クレジット付与されるっぽいので、Omni Flash試したい方はぜひ触ってみてください。右下のウォーターマークもつきません。
Last Seen Users on Sotwe
Penikmat Memek💦
Seen from Indonesia
🎀ELİF🎀
Seen from Turkey
esenyurt travesti 0531=403=16=28
Seen from Turkey
Ensest Cuckold Sex
Seen from Turkey
Penikmat Malam
Seen from Algeria
vcskoleksi
Seen from Indonesia
Bokep hot
Seen from Indonesia
𝙊𝙗𝙚𝙙𝙞𝙚𝙣𝙩 𝙒𝙝𝙤𝙧𝙚𝙨
Seen from France
bannu old man
Seen from United Kingdom
ARMPITS UNIVERSE
Seen from Turkey
Trends for you
Most Popular Users
Elon Musk
@elonmusk
241.5M followers
Barack Obama
@barackobama
119.1M followers
Cristiano Ronaldo
@cristiano
113.7M followers
Donald J. Trump
@realdonaldtrump
111.8M followers
Narendra Modi
@narendramodi
107.2M followers
Rihanna
@rihanna
98.6M followers
NASA
@nasa
92.4M followers
Justin Bieber
@justinbieber
91.7M followers
KATY PERRY
@katyperry
89.6M followers
Taylor Swift
@taylorswift13
83.5M followers
Lady Gaga
@ladygaga
75M followers
Virat Kohli
@imvkohli
72.7M followers
Kim Kardashian
@kimkardashian
70.7M followers
YouTube
@youtube
68.8M followers
Neymar Jr
@neymarjr
65.7M followers
Bill Gates
@billgates
64.9M followers
Selena Gomez
@selenagomez
62.7M followers
The Ellen Show
@theellenshow
62.3M followers
CNN
@cnn
61.8M followers
X
@x
60.8M followers