Made for …

In 10th grade, the words I actually had to learn were sitting in my own textbooks and past exam papers. Every vocabulary app I tried wanted me to study its list instead.

So I built one that starts from my PDF. Drop in a textbook or an exam paper: the extractor pulls out the vocabulary and filters out the common-word noise; an AI “word factory” adds pronunciation, definitions, word roots, collocations and an image cue; and spaced-repetition flashcards drill every word four ways — association, context, spelling, and saying it out loud.

It runs entirely in the browser: no server, no account, no tracking. The deck you build stays on your machine.

One rule I kept from day one: a good tool for learning should not be hidden behind a paywall.

I built it for myself, then made it free so my buddies — and any student — could use it too. It holds a registered software copyright (软件著作权).

I loved a ChatGPT extension that sits beside the web page you are reading. It filled in forms, translated pages, summarised fifty-page documents. Then I tried to set it up for my mum — and she couldn’t use it. In China, the ChatGPT and Gemini browser extensions simply don’t work for her.

DeepSeek does work in China, and it is very good. But DeepSeek’s agent had no browser extension. So I built one, and I made every design decision with one user in mind:

• Her pages stay at home. The agent runs on our own computer; nothing is sent to a third-party browser service.
• It asks before it acts. Reading is free. Clicking, typing and submitting happen one confirmation at a time.
• She will never open a compiler. Installation is three steps — download, run two commands, load the extension — and the setup wizard asks before every change.

Now it reads a whole page, a selection, or a screenshot; and when it is allowed to, it fills forms, translates while keeping the layout, and collects the same information across several tabs.

The experience I loved, made usable by the people around me.

Act 1 · one confident answer

An AI model reviewed my code and said it looked fine. It was wrong, and it sounded completely sure. That was the lesson: one model’s confidence is not evidence.

Act 2 · ask several, blind

PiMoa sends the same question to several models at once. They cannot see each other’s answers, and a final model weighs them. They work inside a read-only sandbox — no internet, no writing to disk, their keys removed — and every claim has to come with a file and a line number. FAILED Version 1 broke on real code: after minutes of work it returned an empty answer.

Act 3 · measure, don’t guess

I did not just raise the timeout. I measured. A 14,000-token task had grown to 251,808 tokens, because every step re-sent the whole conversation. Cited line numbers were off by 130, because the models were counting lines — so PiMoa now prints the real line numbers beside the code, and the models copy them. One model wrote 10,902 characters where the others wrote about 500, so its output is capped. On the same real 1,004-line review, version 2 finished in 22.7 seconds and every cited line number was exact.

Act 4 · a panel can be wrong together

So I added JEV: a judge that answers only closed questions, with a probability. I measured that probability too. Across 212 real judgments, when JEV said it was at least 0.99 sure, it was right 96.1 % of the time; when it said 0.5 – 0.7, only 33.3 %. The rule follows from the data: below 0.7, a human looks. Its blind spots are written on its first page — dates, counting and multi-step reasoning are weak, and given an image it confidently invents an answer, so it refuses images.

Act 5 · keep what works, delete what doesn’t

Its first real job: 3,660 lines of code, five rounds — 10 major and 32 minor problems found, and 7 false alarms (4 of them fake “blockers”) thrown out before they reached me. Then, in October, the numbers showed that an automatic “close the case” step I had designed was not worth what it cost. DELETED I removed it and wrote down why.

An answer is worth what its evidence is worth.

One door for Office files

office-gateway is a single door for Word, Excel and PowerPoint. Behind it, every job goes to the engine that does it best: quick reads and edits to a command-line tool, heavy formula work to a calculation engine, delicate Excel edits through a lossless unpack-and-repack, slides rendered from HTML into real PowerPoint, Word files built directly from their XML format. The engines themselves come from open-source tools and skill packs; what I wrote is the part that decides which hand is safe for which job.

One door for the web

browser-gateway does the same for websites: four engines behind one door — one that clicks and types, 145 adapters that pull structured data from specific sites, read-only diagnostics, and custom scripts. Two lessons live in its comments. On 6 August I found I had described one engine wrongly — it is an installer and does not handle logins — and fixed it that day, because an AI agent believes whatever a tool says about itself. And a live browser is never shared: two agents typing into one tab corrupt each other’s work, so each one gets its own.

The hand I had to build myself

One job no ready-made engine did safely: change one number in a real workbook — full of formulas, charts and hidden sheets — without quietly breaking everything else. An agent asked to do it will often rewrite the whole file; it still opens, and it is still wrong. Sheet Shadow lets the agent edit a shadow copy first, shows the impact, writes only the parts that changed into a new file, and runs a final check for formula errors, damaged structure and changed macros before handing it back.

Routing decides which hand to use. Sheet Shadow is the hand I had to build myself.

By October, three AI assistants on my Mac were using these tools. Then I wanted to reach them from my phone while the Mac was off. Then I wanted assistants from other companies to use them too — including one that cannot install anything at all.

Move A · a shared gateway on my Mac

Each assistant was launching its own copy of every tool, so I built one gateway they all share. One day, six tools started at the same moment and all of them silently hung: macOS gives a background service only 256 file handles. The fix is a single line — 10240 — with a comment beside it saying it must never be deleted.

Move B · a rented cloud computer

To keep working while the Mac sleeps, the same gateway now also runs on a small rented cloud computer, shared by three assistants, behind a private network that only my own devices can join.

Move C · tearing a door down SUPERSEDED

The cloud computer is billed by the minute and sleeps when idle. A permanently open public tunnel counted as activity, so it never slept — about 76 € a month. I tested three ways to get a cheap public entrance; all three were dead ends, and all three are written down so nobody walks them again.

Move D · a free worker on the open internet

One assistant cannot install software, speaks an older version of the protocol, and cannot send login headers. So the tools moved to a Cloudflare Worker that speaks both protocol versions and signs in with OAuth. Today it offers 347 abilities — reading the calendar, sending mail, opening files, writing notes — to 10 AI assistants. The local gateway was not deleted: COLD STANDBY

Moving house carefully is its own skill. One move touched about 40 references across configuration files and scripts. I copied first, checked the copy side by side, switched over in a planned window, watched it for 24 hours, and only then removed the old one — with the rollback practised in advance. It still found a way to break: macOS will not start a background service from an external disk. The fix was to split the layers — the service on the internal disk, the tools wherever they live.

The architecture is a fossil record of things that broke.

Move E · into WeChat

Tools on a server help AI assistants. People live somewhere else — in China, that place is WeChat. Family plans, school notices and the articles worth keeping all arrive there, and none of it ever reached an assistant. So the last move went from inside WeChat to an AI assistant, and back.

Forward it, and it is handled

Forward a message or an article to a bot in WeChat. A small program on an old Mac at home receives it through Tencent’s official bot channel and drops it into an inbox folder. Muse — an AI assistant that lives on its own cloud computer — checks that inbox every minute, puts any dates into the calendar, summarises the article, and writes a reply into an outbox. The program sends the reply back into the same WeChat chat.

Why an old Mac at home? WeChat blocks article requests that come from cloud servers. So the fetching happens on a home connection, and the assistant in the cloud only ever reads the saved text. It can reach exactly one folder on that Mac — nothing else.

Tap “favourite”, and it is filed

The second stream needs no action at all. When I tap “favourite” on an article — a habit I already had — it appears in a knowledge base about two minutes later, sorted into one of 12 folders: by my own rules first, and by a small language model when no rule fits. It now holds about 1,780 items. Forward a WeChat video post to the bot, and it goes straight into the same knowledge base, with a ✅ in the chat.

Ask it, and it answers from my own reading

A knowledge base you cannot ask is just a drawer. So the same corpus gets a personal RAG layer. The full text of every article stays on my own machine as Markdown — the master copy, beyond any app’s lock-in — and a SQLite index recalls candidates two ways: BM25 for exact keywords, multilingual text embeddings for meaning. A small cross-encoder reranker (a qwen3 reranking model) re-sorts the merged list, and the language model I point it at answers with the sources attached. The stack degrades gracefully: if the embedding service is down, keyword search still answers — nothing in the pipeline blocks.

Written down: what does not work

The bot channel does not deliver merged chat-history cards, so the bot politely asks for messages one at a time. Image-only posts have no text, so it asks for a screenshot. Favourited images and files carry no link, so they cannot be filed DEAD END — recorded as one. A reply can never be sent twice: one running copy, a duplicate check, and an atomic claim on every reply. And the honest weak point: the old Mac has to stay on. A plan to move it onto a cloud computer is written down and waiting.

The best interface is the one people already use every day.