DEV Community

Cover image for I translated my 7,000-line AI-coding handbook with 12 parallel AI agents
Dhrupo Nil
Dhrupo Nil

Posted on

I translated my 7,000-line AI-coding handbook with 12 parallel AI agents

Last week I shipped a complete AI-coding handbook in Bangla — 79 terms decoded with everyday analogies, hands-on CLI guides, a 6-chapter story, and a 30-day practice path. The premise was the vocabulary wall: words like "token", "context window", and "harness" sound like spells until someone explains them plainly.

Turns out the vocabulary wall isn't a Bangla problem. It's a beginner problem.

So today the whole book exists in English:

🔗 https://github.com/dhrupo/dictionary-of-ai-coding-english

What's inside (the 60-second version)

📖 Part 1 — The words. 62 terms across 7 sections, each with a daily-life analogy instead of a textbook definition. A model is a calculator that never presses its own buttons. And Goldi 🐠 — a goldfish with a 3-second memory — keeps reminding you that models are stateless too.

🛠️ Part 2 — The tools. 7 hands-on guides: real CLI commands (Claude Code + Codex), AI-friendly folder structure (AGENTS.md, .claude/), bad-prompt → good-prompt rewrites, popular community skills (brainstorming, grill-me, TDD, handoff…), 17 extra terms (RAG, embeddings, temperature, hooks…), token economics, and safety — prompt injection explained with a postman analogy.

📜 Part 3 — The story. Nothing to install: Rafi, a 9th grader, builds his first portfolio site with an AI agent, makes every classic mistake (vague prompts, a hallucinated library, dragging a session deep into the dumb zone), and recovers using the tools from Part 2. You read over his shoulder.

🗓️ The 30-day path. One 5–15 minute mission a day. The first two weeks need nothing but a free AI chat. Plus a myth-busting FAQ, a 79-term index, and copy-paste templates/.

The fun part: the book translated itself the way it teaches you to work

Here's the meta-story. The handbook spends three chapters teaching subagents, handoff artifacts, and verification-before-completion. So when it came time to translate ~7,000 lines of Bangla markdown across 36 files… I used exactly that.

12 parallel AI agents, one session. Each agent got the same conventions block — character names (Goldi, Rafi), recurring section labels, file renames — plus its own slice of the book: three dictionary chapters here, the whole 6-chapter story to a single agent there (one translator = one consistent voice).

Three things made it work:

1. Deterministic rules beat agent judgment. The Bangla book's term headings were ### বাংলা-শব্দ (English Term). The convention: the English heading becomes exactly the parenthetical term. That one rule made every cross-file anchor predictable — agents working on different files in parallel could link into each other's not-yet-written chapters and have the slugs match.

2. A machine-verifiable gate, not vibes. The repo ships a tiny CI link checker (scripts/check-links.py) that validates every markdown link and anchor, GitHub slug quirks included. First run after the agents finished: 44 broken links. Leftover Bangla anchors, one agent guessing #esc-stop where another wrote ### ESC (stopping), a third doubling slugs like #-grill-me-grill-me. Twenty minutes of scripted fixes later: 36 files, every link alive. That checker now runs on every PR.

3. Verify, don't assume. The index agent didn't compute its 79 anchors — it read the actual translated files and extracted the real headings. Zero misses. Meanwhile an asset agent scanned every terminal-GIF source script for Bangla characters and found… none. The GIFs had been authored in English all along. The "re-render 26 GIFs" task evaporated because an agent checked instead of assuming.

If that workflow sounds useful, it's literally what Part 2 and Part 3 of the book teach — subagents, handoffs, verification — just pointed at itself.

It's a community thing

Everything is CC BY 4.0, in both languages, and the two editions cross-link. Issue forms for "I found a mistake" and "suggest a new word" are ready in both repos.

💙 Built on the shoulders of Matt Pocock's Dictionary of AI Coding — Parts 2 and 3 grew far beyond it, but the soul is his.

And the offer from the Bangla launch stands: if you speak another language, steal this structure. The deterministic-heading trick + a link checker means the next translation is mostly an afternoon of agent-wrangling. I'd genuinely love to see a Spanish, Hindi, or Vietnamese edition.

🔗 https://github.com/dhrupo/dictionary-of-ai-coding-english
🇧🇩 Prefer Bangla? https://github.com/dhrupo/dictionary-of-ai-coding-bangla

Top comments (0)