关于本站About

"Japanese Grammar Notes" is a free, bilingual (Chinese + Japanese) JLPT N5 → N2 grammar study set, structured as an 8-week curriculum of 73 lessons covering 400+ grammar points. Each grammar point includes:

• Conjugation rules: how it combines with verbs / adjectives / nouns
• 3+ example sentences: covering daily conversation and exam-style contexts
• Comparisons: side-by-side with easily confused grammar (e.g. は vs. が, たら vs. なら)
• Common pitfalls: mistakes Chinese native speakers tend to make
• Spaced repetition: review checklist for day 0 → 1 → 4 → 7 → 14 → 30

Free, no ads, no tracking.

I started studying Japanese seriously in April 2026, aiming for JLPT N2 in 8 weeks. While preparing, I noticed:

• Chinese-language Japanese grammar material online is either too fragmented (scattered across forums and Q&A sites) or too academic
• Textbooks (Minna no Nihongo, Shin Kanzen Master) leave a gap between "Chinese translation" and "explanation tuned for Chinese speakers" — much of the nuance gets lost
• No single resource covers N5 → N2 completely while explaining from a Chinese-native-speaker perspective

So I decided to take notes as I learn and open-source everything. It started as a personal study aid; publishing it as a website was a side effect.

• GitHub: Ralphbupt
• Email: pengcheng199@gmail.com
• This is a personal project — no team, no commercial intent
• Current JLPT level: studying (progressing N5 → N2 along with this site)
• Preferred channel: GitHub Issues for public tracking; email for private discussion

The lesson structure follows my actual study pace. Example sentences and comparisons reference:

• Textbooks: Minna no Nihongo I/II, Shin Kanzen Master Bunpou N3/N2, Nihongo Sou Matome N3/N2 Bunpou
• Online: Hujiang Japanese, Bunpro, Tofugu, Nihongo Kyoushi no Hiroba
• Past exams: publicly released JLPT N5–N2 grammar questions

Each grammar point is validated through 2–3 weeks of my own study (using the built-in spaced-repetition checklist) before being published.

If you find an example sentence that directly comes from a referenced source without proper attribution, please file an issue — I'll fix it immediately.

The full source code (build scripts, all markdown content) is on GitHub:

https://github.com/Ralphbupt/japanese-grammar

Tech stack:

• marked: Markdown → HTML
• kuroshiro + kuromoji: automatic furigana annotation on kanji
• sharp: OG image generation
• GitHub Pages + Cloudflare Registrar: hosting + custom domain (jpnotes.dev)

Pure static site — no backend, no database, no cookies. Every push to main triggers a GitHub Actions build and deploy.

The 🤖 button on every lesson page opens a side panel; the AI has already read the lesson, so just ask. Click "Explain" after any example sentence for a short grammar breakdown with translation and pitfalls, or select any text and click "💬 Ask AI" to put it in the input box and ask your own question.

It works out of the box: by default it uses a free model this site provides (OpenAI's open-weight gpt-oss-120b, running on Cloudflare), with no sign-up and no key, 50 questions per reader per day. For no daily limit or a stronger model, use your own API key. Groq is the easy choice: free, no card, answers in about a second.

1. Open console.groq.com/keys and sign in with Google or GitHub.
2. Click Create API Key, give it a name and copy the key (it starts with gsk_).
3. Back on any lesson page, click 🤖, click "Use your own API key", pick "Groq", paste the key and click "Connect and start".

Do not pick xAI Grok by mistake; that is Elon Musk’s paid model, one letter apart.

Other free options

• Google Gemini: the Flash models have a free tier; create a key at aistudio.google.com.
• OpenRouter: one key for every model; models ending in :free cost nothing (50 requests a day).
• Fully local: install Ollama, run ollama pull qwen3:14b, then start it as the settings hint says with OLLAMA_ORIGINS=https://jpnotes.dev ollama serve. A local 14B model takes about 30 s per question; the Quick setting skips its thinking phase.

Troubleshooting

• 429 / rate limit: Groq’s free tier allows 8,000 tokens a minute and a whole-lesson question uses 5–6k, so roughly one question per minute. The 📎 selector above the input defaults to "follow my reading", which sends only the grammar point on screen and uses far fewer tokens; you can also pick a section by hand.
• Failed to fetch: the provider blocks browser calls, or local Ollama was started without OLLAMA_ORIGINS. Switch to a provider that supports browser calls, such as Groq or OpenRouter.
• Switching providers or models: click the model name in the panel header; every provider you have configured is in that menu, along with the three thinking-depth settings.

Your own key never leaves your browser and never passes through this site. See the privacy section below.

The 🤖 "Ask AI" side panel on lesson pages sends the current lesson plus your question to a large language model. Its design rule: this site never receives your key and never stores your conversation.

• The default free model goes through this site's relay. The "Free" option uses a Cloudflare Worker (ai.jpnotes.dev) that passes your question and the lesson text to a model on Cloudflare Workers AI and streams the answer back. It does not log questions or answers and has request logging turned off; to enforce the daily limit it keeps only a salted hash of your IP, counted per day and deleted after two days. Beyond that it only counts which button each request came from (Explain, quick prompts, …) and how much quota it used — no text, no IP. The relay source is public too. To bypass the site entirely, use your own key.
• Your own key never touches this site. The site itself is static HTML on GitHub Pages. Your API key exists only in your browser on your device (localStorage by default; tick "keep for this tab only" and it is cleared when the tab closes).
• With your own key, requests go straight to the provider. When you ask, your browser sends the lesson text and your question straight to the provider you picked (Anthropic, OpenAI, DeepSeek, Qwen… or Ollama on your own machine). Nothing sits in between.
• A browser-enforced allowlist. Every page ships a Content-Security-Policy connect-src rule that only permits connections to this site, Google Analytics, the preset provider hosts, and localhost / 127.0.0.1. Even if a script on the page were tampered with, the browser would refuse to send data anywhere else.
• No plain-text keys. Settings accept only https:// or local URLs, and changing a provider's host to a non-official domain shows a prominent warning.
• Verify it yourself. Open DevTools (F12) → Network and ask a question: with your own key, the only requests you will see go to that provider's domain. The panel's entire source is one file; reviews welcome.
• Fully local option. To keep everything on your machine, run a local model with Ollama or LM Studio; the browser then only talks to localhost.

The free model costs you nothing; with your own key, billing is between you and the provider, and this site takes no part and no cut.

Japanese Daily News is our companion podcast — slow Japanese news with English commentary, designed for learners. The website offers sentence-by-sentence listening, transcripts with furigana, while episodes are also available on Spotify / Apple Podcasts.

Found a typo? Awkward example? Unclear explanation?

• GitHub Issues: open an issue — recommended, publicly tracked
• GitHub PR: submit a fix directly; additions to examples and comparison notes especially welcome
• Email: pengcheng199@gmail.com

I respond to every piece of feedback.

Content is licensed under CC BY 4.0:

• ✅ Free to share, quote, adapt, use for teaching (including commercial teaching)
• ✅ Just attribute the source (link to jpnotes.dev)
• ❌ Please don't mirror the entire site — full mirrors dilute Google's index and hurt both the original and the mirror. Quote any snippet or chapter freely, but please don't bulk-copy the entire grammar point catalog