https://drpwchen.com/en/tags/claude/Recent content in Claude on Po-Wei Chen, MD — Physiatrist & Builderhttps://drpwchen.com/og-default.pnghttps://drpwchen.com/og-default.pngHugoenTue, 14 Jul 2026 21:40:00 +0800https://drpwchen.com/en/posts/paper-fetch/Tue, 14 Jul 2026 08:15:00 +0800https://drpwchen.com/en/posts/paper-fetch/To let AI read papers for you, you first need the full text. And what usually gets you stuck isn't the AI—it's permissions, publishers, and the library. paper-fetch uses a step-by-step route from open access, to official publisher APIs, to your own hospital library to turn a DOI into a full-text PDF, without taking any piracy routes.https://drpwchen.com/en/posts/my-data-my-benchmark/Mon, 13 Jul 2026 16:31:00 +0800https://drpwchen.com/en/posts/my-data-my-benchmark/MediaTek's official benchmark already proves Breeze-ASR-25 beats Whisper, but it isn't testing your task. Using a 180,000-word glossary built from hundreds of textbooks as a yardstick, I found it captures 49% more real vocabulary and runs faster. Another finding not in any model card: the time alignment of the Taiwanese version, Breeze-ASR-26, has regressed. Default decoding only segments once every 30 seconds, which outright breaks when making subtitles or aligning with slides, but enabling word-level timestamps saves it. The testing method is open-sourced.https://drpwchen.com/en/posts/paper-tools/Tue, 07 Jul 2026 09:30:00 +0800https://drpwchen.com/en/posts/paper-tools/I took the entire workflow of a journal club critical appraisal and wrote it into two Claude Code skills. I even wrote two deterministic little scripts to hand over the calculation of evidence certainty and logical flaws to code, stopping the AI from bullshitting based on overall impressions. It's already open-sourced on GitHub.https://drpwchen.com/en/posts/ai-token-4-budget-dashboard/Mon, 29 Jun 2026 06:30:00 +0800https://drpwchen.com/en/posts/ai-token-4-budget-dashboard/The final part of the saving tokens series: not everything needs the smartest brain, and even for reading images, the cheapest model is often enough. Toss noisy chores to a clone to keep your desk clean, then install a dashboard, or even let the AI watch its own usage and pace the budget just right.https://drpwchen.com/en/posts/ai-token-2-script-vs-llm/Mon, 29 Jun 2026 06:30:00 +0800https://drpwchen.com/en/posts/ai-token-2-script-vs-llm/Some jobs simply shouldn't be given to AI. Clear-cut, repetitive tasks belong to hardcoded scripts—they're fast, accurate, and practically free. Vague, judgment-heavy tasks are what you boot up AI for. Here's how I read my medical textbook PDFs as an example of making this tradeoff.https://drpwchen.com/en/posts/ai-token-3-clean-desk/Mon, 29 Jun 2026 06:30:00 +0800https://drpwchen.com/en/posts/ai-token-3-clean-desk/There's a cost you pay with every sentence: the config file residing right in front of the AI. Keep it lean, layer the memory, make good use of prompt caching to bill repeated content at one-tenth the price, and throw away the noise of successful commands directly. A clean desk doesn't just save money; the AI answers more accurately, too.https://drpwchen.com/en/posts/ai-token-1-chat-gets-dumber/Mon, 29 Jun 2026 06:30:00 +0800https://drpwchen.com/en/posts/ai-token-1-chat-gets-dumber/AI has no memory. With every reply, it rereads the entire conversation from the beginning, which is why longer chats get slower, more expensive, and dumber. This post explains the mechanism using a taxi meter analogy, and gives everyone who chats with AI three ways to save that you can use today.https://drpwchen.com/en/posts/rehab-course-radar/Wed, 24 Jun 2026 07:30:00 +0800https://drpwchen.com/en/posts/rehab-course-radar/I built a website, courses.drpwchen.com, that automatically aggregates courses and announcements from 17 rehabilitation-related societies and institutions. It updates every 3 hours, supports RSS or ntfy notifications, and is free with no login required.https://drpwchen.com/en/posts/vault-graph-rag/Mon, 22 Jun 2026 15:00:00 +0800https://drpwchen.com/en/posts/vault-graph-rag/I threw a new paper on SQL-RAG at Claude to verify it myself: useless for a redundant textbook library, but a great fit for finding "related notes" in a sparse Obsidian medical note vault. After swapping the ranking to personalized PageRank, the proportion of correctly finding related notes jumped from 34% to 84%, and I put it straight into production.