<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: XM Yang</title>
    <description>The latest articles on DEV Community by XM Yang (@xmyang0223).</description>
    <link>https://dev.to/xmyang0223</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4210078%2Fad654f88-b9e6-4bbc-8837-20311c4f5858.png</url>
      <title>DEV Community: XM Yang</title>
      <link>https://dev.to/xmyang0223</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/xmyang0223"/>
    <language>en</language>
    <item>
      <title>Pay Once, Own Forever: 22 Tools That Replace a Subscription</title>
      <dc:creator>XM Yang</dc:creator>
      <pubDate>Tue, 03 Nov 2026 17:11:21 +0000</pubDate>
      <link>https://dev.to/xmyang0223/pay-once-own-forever-22-tools-that-replace-a-subscription-25k</link>
      <guid>https://dev.to/xmyang0223/pay-once-own-forever-22-tools-that-replace-a-subscription-25k</guid>
      <description>&lt;p&gt;Every tool wants $10–20 a month now, forever. So I keep coming back to one list: Pay Once, Own Forever on Tools Hub — &lt;a href="https://tools-hub.sweetai.work/pay-once" rel="noopener noreferrer"&gt;https://tools-hub.sweetai.work/pay-once&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;It collects 22 hand-verified tools you buy once (or use free), and each entry is shown next to the subscription tool it replaces, that subscription's yearly cost, and the five-year total. Seeing the five-year number next to a one-time price is what makes the trade-off obvious.&lt;/p&gt;

&lt;p&gt;The list lives inside a much bigger directory: Tools Hub (&lt;a href="https://tools-hub.sweetai.work" rel="noopener noreferrer"&gt;https://tools-hub.sweetai.work&lt;/a&gt;) catalogues 10,325 tools across 11 categories, with best-of guides and rankings, so after the pay-once page the same site is where I check what else exists in a category before paying for anything monthly.&lt;/p&gt;

&lt;p&gt;Not everything should be bought outright — but for the tools you will still be using in five years, this page is a good first stop.&lt;/p&gt;

</description>
      <category>softwaredevelopment</category>
      <category>productivity</category>
      <category>tools</category>
      <category>saas</category>
    </item>
    <item>
      <title>How to Compress a PDF Without Uploading It Anywhere</title>
      <dc:creator>XM Yang</dc:creator>
      <pubDate>Tue, 03 Nov 2026 17:07:36 +0000</pubDate>
      <link>https://dev.to/xmyang0223/how-to-compress-a-pdf-without-uploading-it-anywhere-1m8g</link>
      <guid>https://dev.to/xmyang0223/how-to-compress-a-pdf-without-uploading-it-anywhere-1m8g</guid>
      <description>&lt;p&gt;The email said my PDF was too large to send. The usual fix — a free online compressor — wanted me to upload a signed contract to a server I know nothing about, then wait in a queue behind a daily limit.&lt;/p&gt;

&lt;p&gt;So I went looking for a compressor that never takes the file in the first place, and found Localcraft's Compress PDF: &lt;a href="https://tools.sweetai.work/compress-pdf/" rel="noopener noreferrer"&gt;https://tools.sweetai.work/compress-pdf/&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;It runs entirely in your browser. You drop the PDF in, it compresses on your own device, and you download the result. No upload, no sign-up, no watermark, and no daily cap — which also means it works the same on the tenth file as on the first.&lt;/p&gt;

&lt;p&gt;That local processing is the whole point of the site behind it: Localcraft (&lt;a href="https://tools.sweetai.work" rel="noopener noreferrer"&gt;https://tools.sweetai.work&lt;/a&gt;) is a collection of 112 free tools — 44 of them PDF tools, plus image, AI and developer tools — and they all work the same way, in the browser, without your files leaving the machine.&lt;/p&gt;

&lt;p&gt;If the file is a contract, an ID scan, or anything you would not email to a stranger, compressing it locally is the version that makes sense.&lt;/p&gt;

</description>
      <category>productivity</category>
      <category>privacy</category>
      <category>webdev</category>
    </item>
    <item>
      <title>61.5% of Shopify Stores Still Run Script Tags That Die on March 1, 2027 — I Scanned 253 of Them</title>
      <dc:creator>XM Yang</dc:creator>
      <pubDate>Mon, 02 Nov 2026 08:17:52 +0000</pubDate>
      <link>https://dev.to/xmyang0223/615-of-shopify-stores-still-run-script-tags-that-die-on-march-1-2027-i-scanned-253-of-them-2283</link>
      <guid>https://dev.to/xmyang0223/615-of-shopify-stores-still-run-script-tags-that-die-on-march-1-2027-i-scanned-253-of-them-2283</guid>
      <description>&lt;p&gt;Shopify has set a hard date: on &lt;strong&gt;March 1, 2027&lt;/strong&gt;, script tags stop running. Not deprecated-with-a-warning, not phased out quietly — stopped. Every tracking pixel, review widget, and app snippet that still reaches the storefront through a script tag has to move to Web Pixels, App Embeds, or another supported route before that day, or it simply goes dark.&lt;/p&gt;

&lt;p&gt;That date is about sixteen months out, which in store-maintenance time is basically tomorrow. So I got curious: how ready are stores, actually? Not in survey answers — in their live page source. I scanned 253 publicly listed Shopify storefronts on November 2, 2026, counting what is really loaded on their home, product, and collection pages.&lt;/p&gt;

&lt;h2&gt;
  
  
  Finding 1: 155 of 252 stores (61.5%) still run at least one tag that dies on deadline day
&lt;/h2&gt;

&lt;p&gt;Of the 252 stores with complete script-tag data, &lt;strong&gt;155 (61.5%)&lt;/strong&gt; run at least one script tag that will stop working on March 1, 2027. The median store carries just 1.0 such tags, but the tail is long: 46 of 252 stores run 4–7 retiring tags, 13 run 8–12, and one runs 13 or more. Each of those tags is a piece of storefront functionality — analytics, ads, reviews, chat, consent — with a scheduled outage attached. One tag is an afternoon of migration work. Ten tags is a project nobody has put on the roadmap yet.&lt;/p&gt;

&lt;h2&gt;
  
  
  Finding 2: 118 of 251 stores (47.0%) load the same script twice on a single page — the worst loaded one URL 24 times
&lt;/h2&gt;

&lt;p&gt;This one surprised me more than the headline number. &lt;strong&gt;118 of 251 stores (47.0%)&lt;/strong&gt; load the &lt;em&gt;same script URL two or more times on a single page&lt;/em&gt;. The worst single case loaded one URL &lt;strong&gt;24 times&lt;/strong&gt; on one page. Duplicate loads are what happens when a snippet gets pasted into a theme, then added again by an app, then a third time by a tag manager — and nobody audits the pile. They waste bandwidth on every visit, and they make migration harder: before you can move a tag, you first have to work out which of its three copies is the real one. The containers looked cleaner — only 13 of 251 stores (5.2%) run more than one GA4 destination, and just 1 of 251 runs more than one GTM container.&lt;/p&gt;

&lt;h2&gt;
  
  
  Finding 3: 189 of 550 retiring tags (34.4%) have no automatic migration path
&lt;/h2&gt;

&lt;p&gt;Across the sample I catalogued &lt;strong&gt;550 retiring tags&lt;/strong&gt; and classified where each one can go. The good news: 329 (59.8%) map to App Embeds and 28 (5.1%) to Web Pixels, both supported routes. The bad news: &lt;strong&gt;189 tags (34.4%) have no automatic migration path at all&lt;/strong&gt; and need manual review — someone has to open each one, work out what it does, and decide how to replace it. That is the number I would plan around. The automatic share of this migration is the easy share; a third of it is hand work, and hand work is exactly what gets deferred until the deadline is close enough to hurt. One more quiet finding in the same scan: 20 of 251 stores (8.0%) expose a plaintext token-type marker in their public page source. I recorded counts only, never token values, but it is a reminder that storefront source accumulates things nobody re-reads.&lt;/p&gt;

&lt;h2&gt;
  
  
  Method, honestly
&lt;/h2&gt;

&lt;p&gt;I attempted 300 storefronts and successfully scanned 253. The sample is &lt;strong&gt;publicly listed Shopify stores that were reachable during the scan week&lt;/strong&gt; — drawn from Shopify's official examples and showcase listings, Theme Store demo and featured store links, and stores self-reported in public Shopify Community topics. Failed scans (unreachable, not on Shopify, password-protected) are excluded, and every figure above carries its own denominator (251–252, depending on field completeness) for exactly that reason. This is not a random sample of all Shopify merchants — it describes these 253 stores, scanned on one day, from their public page source. For token flags I stored only a yes/no marker and a type word, never a value. Store-level results are anonymized.&lt;/p&gt;

&lt;h2&gt;
  
  
  The full data
&lt;/h2&gt;

&lt;p&gt;The full report — all six metrics, the per-store tag distribution, migration-path breakdown, and the complete method notes — is here:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;&lt;a href="https://tagaudit.sweetai.work/report/2026-script-tag-countdown/" rel="noopener noreferrer"&gt;The 2027 Script-Tag Countdown: Storefront Tracking Health Report&lt;/a&gt;&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Everything in it is free to quote with figure, denominator, and a link back. And if you run a Shopify store yourself, you can check your own storefront for free and see which of your tags are on the retiring list — that takes about a minute and beats finding out on March 1, 2027.&lt;/p&gt;

&lt;p&gt;I'm sharing the raw distribution tables with anyone who wants to dig in, so if you want a specific cut of the data, ask in the comments. And a question for the Shopify devs and app builders here: does the 61.5% match what you're seeing on the stores you work with — or is your corner of the ecosystem further ahead than this sample?&lt;/p&gt;

</description>
      <category>shopify</category>
      <category>webdev</category>
      <category>analytics</category>
      <category>javascript</category>
    </item>
    <item>
      <title>Can you open Stripe from mainland China? I checked the official list so you don't have to</title>
      <dc:creator>XM Yang</dc:creator>
      <pubDate>Mon, 02 Nov 2026 01:24:14 +0000</pubDate>
      <link>https://dev.to/xmyang0223/can-you-open-stripe-from-mainland-china-i-checked-the-official-list-so-you-dont-have-to-28oi</link>
      <guid>https://dev.to/xmyang0223/can-you-open-stripe-from-mainland-china-i-checked-the-official-list-so-you-dont-have-to-28oi</guid>
      <description>&lt;p&gt;Short answer: no, not directly. Stripe's official availability list (stripe.com/global) does not include mainland China. Hong Kong is listed separately as "Hong Kong SAR" — it does not cover a mainland entity. Stripe onboards by business country, so mainland China is not even selectable in the signup flow. Prices and list status checked 2026-10-30; if anything here conflicts with Stripe's current pages, the official pages prevail.&lt;/p&gt;

&lt;p&gt;That leaves three real paths:&lt;/p&gt;

&lt;h2&gt;
  
  
  1. A US company via Stripe Atlas — $500, one time
&lt;/h2&gt;

&lt;p&gt;Atlas forms a Delaware C-corp or LLC online: state filing, EIN, founder equity paperwork, and the first year of registered-agent service included. After year one, the registered agent runs about $100/year, plus Delaware taxes and US filings. Only worth it once you have steady overseas revenue — $500 is the entry ticket, not the total cost.&lt;/p&gt;

&lt;h2&gt;
  
  
  2. A Hong Kong company
&lt;/h2&gt;

&lt;p&gt;Hong Kong is on Stripe's list, so a Hong Kong limited company can open a Hong Kong Stripe account. But incorporation, bank onboarding (getting stricter), and annual audit and tax filing are ongoing costs. Budget "can register" and "can maintain cheaply" as two different numbers.&lt;/p&gt;

&lt;h2&gt;
  
  
  3. Skip Stripe — let a platform collect for you
&lt;/h2&gt;

&lt;p&gt;For digital products and client work, this is the right default until volume justifies an entity. On a $29 sale, using each platform's published rates:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Channel&lt;/th&gt;
&lt;th&gt;You keep on $29&lt;/th&gt;
&lt;th&gt;How money reaches you&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Gumroad (direct)&lt;/td&gt;
&lt;td&gt;$24.46&lt;/td&gt;
&lt;td&gt;Payout to PayPal once you pass $100&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Payhip (Free plan)&lt;/td&gt;
&lt;td&gt;$26.41&lt;/td&gt;
&lt;td&gt;Payout to PayPal / Stripe&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;PayPal direct&lt;/td&gt;
&lt;td&gt;Varies by account rate&lt;/td&gt;
&lt;td&gt;Wire to a Chinese bank: $35 per wire, $150 minimum&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Worked examples are arithmetic on official rates, not an earnings promise; Payhip sales also carry card processing (roughly 2.9% + $0.30), so the real take is slightly lower.&lt;/p&gt;

&lt;h2&gt;
  
  
  How to choose
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;No sales yet:&lt;/strong&gt; path 3. Validate that strangers pay you first; the fee difference on small volume is a few dollars.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Steady volume, subscriptions/SaaS:&lt;/strong&gt; run the Atlas math properly, including yearly maintenance.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Already have a Hong Kong entity:&lt;/strong&gt; path 2 as a bonus — don't maintain a company just for Stripe.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The full comparison, including why borrowed-address "workarounds" fail Stripe's verification, is in the free bilingual guide: &lt;a href="https://go.sweetai.work/en/guides/stripe-china/" rel="noopener noreferrer"&gt;https://go.sweetai.work/en/guides/stripe-china/&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Which path did you take — and what surprised you about the real cost?&lt;/p&gt;

</description>
      <category>stripe</category>
      <category>payments</category>
      <category>indiehackers</category>
      <category>startup</category>
    </item>
    <item>
      <title>Found a handy prompt library</title>
      <dc:creator>XM Yang</dc:creator>
      <pubDate>Sat, 31 Oct 2026 06:10:01 +0000</pubDate>
      <link>https://dev.to/xmyang0223/found-a-handy-prompt-library-2889</link>
      <guid>https://dev.to/xmyang0223/found-a-handy-prompt-library-2889</guid>
      <description>&lt;p&gt;Found a handy prompt library — this one turns a topic into a short newsletter explainer, and you can fill in the variables and watch the prompt update live: &lt;a href="https://prompts.sweetai.work/p/a-short-explainer-on-how-tides-work-for-a-small-business-audience-newsletter-edition-everyday-writing" rel="noopener noreferrer"&gt;https://prompts.sweetai.work/p/a-short-explainer-on-how-tides-work-for-a-small-business-audience-newsletter-edition-everyday-writing&lt;/a&gt;  Its image section alone has 137 ready-to-use prompts: &lt;a href="https://prompts.sweetai.work/category/image-generation/" rel="noopener noreferrer"&gt;https://prompts.sweetai.work/category/image-generation/&lt;/a&gt;&lt;/p&gt;

</description>
    </item>
    <item>
      <title>Found a handy prompt library</title>
      <dc:creator>XM Yang</dc:creator>
      <pubDate>Fri, 30 Oct 2026 11:20:27 +0000</pubDate>
      <link>https://dev.to/xmyang0223/found-a-handy-prompt-library-3c2f</link>
      <guid>https://dev.to/xmyang0223/found-a-handy-prompt-library-3c2f</guid>
      <description>&lt;p&gt;Found a handy prompt library — this one turns a topic into a short newsletter explainer, and you can fill in the variables and watch the prompt update live: &lt;a href="https://prompts.sweetai.work/p/a-short-explainer-on-how-tides-work-for-a-small-business-audience-newsletter-edition-everyday-writing" rel="noopener noreferrer"&gt;https://prompts.sweetai.work/p/a-short-explainer-on-how-tides-work-for-a-small-business-audience-newsletter-edition-everyday-writing&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Its image section alone has 137 ready-to-use prompts: &lt;a href="https://prompts.sweetai.work/category/image-generation/" rel="noopener noreferrer"&gt;https://prompts.sweetai.work/category/image-generation/&lt;/a&gt;&lt;/p&gt;

</description>
    </item>
    <item>
      <title>Sell, hold, or buy the dip?</title>
      <dc:creator>XM Yang</dc:creator>
      <pubDate>Fri, 30 Oct 2026 11:00:01 +0000</pubDate>
      <link>https://dev.to/xmyang0223/sell-hold-or-buy-the-dip-p83</link>
      <guid>https://dev.to/xmyang0223/sell-hold-or-buy-the-dip-p83</guid>
      <description>&lt;p&gt;Found a neat little tool for when you're stuck on a decision — it lays out another point of view instead of just telling you what to do. Today's one felt timely: after a red day in stocks and crypto, do you sell, hold, or buy the dip? &lt;a href="https://decide.sweetai.work/q/3e66e520-e9e8-42e9-8c54-70ca2e9b81a3" rel="noopener noreferrer"&gt;https://decide.sweetai.work/q/3e66e520-e9e8-42e9-8c54-70ca2e9b81a3&lt;/a&gt;&lt;/p&gt;

</description>
    </item>
    <item>
      <title>The LiteLLM incident check you shouldn't run with Python — and the scanner I built for it</title>
      <dc:creator>XM Yang</dc:creator>
      <pubDate>Wed, 28 Oct 2026 12:19:27 +0000</pubDate>
      <link>https://dev.to/xmyang0223/the-litellm-incident-check-you-shouldnt-run-with-python-and-the-scanner-i-built-for-it-5gc6</link>
      <guid>https://dev.to/xmyang0223/the-litellm-incident-check-you-shouldnt-run-with-python-and-the-scanner-i-built-for-it-5gc6</guid>
      <description>&lt;p&gt;On Oct 21, 2026, LiteLLM disclosed that its v1.107.0 release had been compromised and published to PyPI, Docker Hub, and GHCR before being contained.  The part of the official response steps that stuck with me: in an affected environment, &lt;strong&gt;any Python start re-runs the implant&lt;/strong&gt;. So the vendor tells you not to begin with &lt;code&gt;pip install&lt;/code&gt; or &lt;code&gt;pip show&lt;/code&gt; on a suspect box. The first-pass checks are file and metadata checks, done from the shell, by hand:  - Is &lt;code&gt;litellm&lt;/code&gt; pinned to exactly 1.107.0 anywhere — or does the version spec &lt;em&gt;admit&lt;/em&gt; 1.107.0 (unpinned, &lt;code&gt;&amp;gt;=&lt;/code&gt;, &lt;code&gt;~=&lt;/code&gt;, &lt;code&gt;^&lt;/code&gt;)? - What did the lockfiles actually resolve to? - Do any Dockerfiles or compose files reference the compromised tags (&lt;code&gt;v1.107.0&lt;/code&gt;, &lt;code&gt;latest&lt;/code&gt;, &lt;code&gt;main-latest&lt;/code&gt;) or the published bad image digest? - Is there a &lt;code&gt;litellm-1.107.0.dist-info&lt;/code&gt; in any virtualenv on the box? - Are the implant marker files there — &lt;code&gt;cook.pth&lt;/code&gt;, &lt;code&gt;cook.py&lt;/code&gt;, &lt;code&gt;remote_code_loader.lock&lt;/code&gt;?  Doing that by hand across a few projects and a folder of virtualenvs is exactly the kind of boring, error-prone pass a small tool should do. So I built one.  ## litellm-check&lt;br&gt;
&lt;br&gt;
  &lt;code&gt;bash pip install litellm-check litellm-check scan ~/src/my-service&lt;/code&gt;&lt;br&gt;
&lt;br&gt;
  Zero third-party dependencies, Python 3.11+. It reads files and package metadata only — it never executes Python (or pip) inside a scanned environment, which is the whole point: the check itself must be safe to run on a suspect machine, ideally with a clean interpreter.  A few design choices worth noting:  &lt;strong&gt;Specs are evaluated, not grepped.&lt;/strong&gt; A pin of exactly &lt;code&gt;1.107.0&lt;/code&gt; is critical. A spec that admits it is high — exposure surface, not proof of installation, because whether it actually resolved to 1.107.0 depends on installs inside the announced windows (PyPI 11:39:51–13:23 UTC on Oct 21; Docker Hub until 14:20 UTC; GHCR until 13:51 UTC). A spec that excludes it is reported as info, so a reviewer can see the coverage instead of a silent pass.  &lt;strong&gt;Every rule traces to the announcement.&lt;/strong&gt; The 13 rules (LL-001–LL-013) are derived solely from LiteLLM's official incident post — no third-party IOC lists. Each finding in the report carries the file:line evidence, the incident basis, and the remediation (delete the implant files from the shell, kill the daemon holding the lock file, rebuild pinned to &lt;code&gt;litellm==1.106.2&lt;/code&gt;, rotate reachable secrets, check CloudTrail).  &lt;strong&gt;The limits are printed on every report.&lt;/strong&gt; A clean scan means "none of the known indicators were found in this tree" — nothing more. Environments rebuilt after exposure, images pulled without any definition file, installs recorded only in CI logs: none of that leaves anything for a file scanner to find. The tool collects evidence; it does not certify anything.  The repo ships a deliberately exposed demo project; scanning it produces 16 findings (10 critical, 4 high, 2 info) so you can see the report shape before pointing it at anything real: &lt;a href="https://github.com/dylanXM/litellm-check" rel="noopener noreferrer"&gt;https://github.com/dylanXM/litellm-check&lt;/a&gt;  ## The sibling tool: DeepSeek Exposure Scan  Same rules-as-data engine shape, different problem. DHS Exclusion and Removal Order #2026-01 (published in the Federal Register on Oct 20) directs federal civilian agencies to remove DeepSeek products — including third-party products built using DeepSeek implementations — and FAR 52.204-30 flows that down to contractors, who now get asked by customers: &lt;em&gt;what model components are inside what you ship?&lt;/em&gt;  &lt;code&gt;deepseek-scan&lt;/code&gt; (&lt;a href="https://github.com/dylanXM/deepseek-scanner" rel="noopener noreferrer"&gt;https://github.com/dylanXM/deepseek-scanner&lt;/a&gt;) does the first evidence pass for that question: hosted API endpoints, SDK imports, model ids, dependency manifests and lockfiles, container model pulls, and local weight files — written up as a Markdown or AI-BOM-style JSON report, with the same honesty about what it cannot detect (a renamed, distilled model leaves no textual trace).&lt;br&gt;
&lt;br&gt;
  &lt;code&gt;bash pip install deepseek-scanner&lt;/code&gt;&lt;br&gt;
&lt;br&gt;
  Both tools were recently reviewed and merged into the awesome-ai-security-tools list (in its watchlist section — they're new, and the maintainer is right that adoption signals come first). If you try either one, the feedback I value most is false positives and missing rules: both rule sets are single JSON files, so a new detection is a data change, not a code change.  &lt;em&gt;Disclosure: I'm the author of both tools. They're Apache-2.0 and not affiliated with BerriAI/LiteLLM or DeepSeek.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>python</category>
      <category>security</category>
      <category>opensource</category>
    </item>
    <item>
      <title>How indie makers in China actually get paid by overseas customers (with the real fee math)</title>
      <dc:creator>XM Yang</dc:creator>
      <pubDate>Tue, 27 Oct 2026 07:35:07 +0000</pubDate>
      <link>https://dev.to/xmyang0223/how-indie-makers-in-china-actually-get-paid-by-overseas-customers-with-the-real-fee-math-4mdo</link>
      <guid>https://dev.to/xmyang0223/how-indie-makers-in-china-actually-get-paid-by-overseas-customers-with-the-real-fee-math-4mdo</guid>
      <description>&lt;p&gt;If you build from China and sell to overseas customers, the product is the easy part. The hard part is the money plumbing: which channel actually lands money in a Chinese bank account, and how much of each sale survives the trip.&lt;/p&gt;

&lt;p&gt;I spent weeks verifying the official fee pages and running the numbers. Here is the short version.&lt;/p&gt;

&lt;h2&gt;
  
  
  Route 1: PayPal direct, then wire to a Chinese bank
&lt;/h2&gt;

&lt;p&gt;Best for direct client work. The catch is the withdrawal: PayPal China wires cost a &lt;strong&gt;flat $35 per withdrawal, minimum $150, 3–7 business days&lt;/strong&gt;. That flat fee punishes small balances:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;You withdraw&lt;/th&gt;
&lt;th&gt;Flat fee eats&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;$150&lt;/td&gt;
&lt;td&gt;23.3%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;$350&lt;/td&gt;
&lt;td&gt;10%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;$700&lt;/td&gt;
&lt;td&gt;5%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;$1,000&lt;/td&gt;
&lt;td&gt;3.5%&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Rule of thumb: let the balance accumulate to at least $700 before wiring, unless cash flow forces your hand.&lt;/p&gt;

&lt;h2&gt;
  
  
  Route 2: Let a platform collect for you (Gumroad / Payhip)
&lt;/h2&gt;

&lt;p&gt;Best for digital products — templates, ebooks, toolkits. The platform takes its cut per sale but handles cards, tax plumbing and payout.&lt;/p&gt;

&lt;p&gt;What a sale is actually worth after the platform fee (card processing included, using each platform's published rates):&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Price&lt;/th&gt;
&lt;th&gt;Gumroad direct&lt;/th&gt;
&lt;th&gt;Gumroad Discover&lt;/th&gt;
&lt;th&gt;Payhip free plan&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;$10&lt;/td&gt;
&lt;td&gt;$7.91&lt;/td&gt;
&lt;td&gt;$7.00&lt;/td&gt;
&lt;td&gt;$8.91&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;$29&lt;/td&gt;
&lt;td&gt;$24.46&lt;/td&gt;
&lt;td&gt;$20.30&lt;/td&gt;
&lt;td&gt;$26.41&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;$99&lt;/td&gt;
&lt;td&gt;$85.43&lt;/td&gt;
&lt;td&gt;$69.30&lt;/td&gt;
&lt;td&gt;$90.88&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Two lessons: Discover/marketplace traffic is expensive (30% on Gumroad), and at small price points Payhip's 5% free plan beats Gumroad's 10% + $0.50 surprisingly often.&lt;/p&gt;

&lt;h2&gt;
  
  
  Route 3: Marketplaces pay you on their schedule
&lt;/h2&gt;

&lt;p&gt;Etsy, app stores and freelance platforms (project fee 10% or $5 minimum on Freelancer, milestone escrow) remove the collection problem entirely — you trade margin and control for a built-in payout rail. Fine for validation, painful as a permanent home.&lt;/p&gt;

&lt;h2&gt;
  
  
  The stack I recommend
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Client work&lt;/strong&gt; → PayPal invoice, batch withdrawals at $700+&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Digital products&lt;/strong&gt; → Payhip or Gumroad as the cashier, drive your own traffic (never rely on Discover at 30%)&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Validation&lt;/strong&gt; → whatever marketplace already has your buyers&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;I turned the full research into a free bilingual playbook site — payment maps, platform thresholds, and 12 field-by-field playbooks (e-commerce, content, digital products, freelancing, POD, affiliate and more), each with a worked profit example and a first-week checklist: &lt;strong&gt;&lt;a href="https://go.sweetai.work/en/" rel="noopener noreferrer"&gt;go.sweetai.work&lt;/a&gt;&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;If you're selling overseas from China, which channel burned you the most? I'd genuinely like to add the missing cases to the playbook.&lt;/p&gt;

</description>
      <category>indiehackers</category>
      <category>startup</category>
      <category>payments</category>
      <category>showdev</category>
    </item>
    <item>
      <title>"AI agent" is becoming meaningless — here's the test I use</title>
      <dc:creator>XM Yang</dc:creator>
      <pubDate>Tue, 27 Oct 2026 06:30:37 +0000</pubDate>
      <link>https://dev.to/xmyang0223/ai-agent-is-becoming-meaningless-heres-the-test-i-use-e84</link>
      <guid>https://dev.to/xmyang0223/ai-agent-is-becoming-meaningless-heres-the-test-i-use-e84</guid>
      <description>&lt;p&gt;Half the products calling themselves "AI agents" are chatbots with a new landing page. When I built the Personal Agents section on Tools Hub, I needed a definition that excludes them.&lt;/p&gt;

&lt;p&gt;The test: does it do work in the background, on its own computer, with its own identity — while you're not in the chat? If it only replies when you type, it's a companion or a chatbot, not a personal agent.&lt;/p&gt;

&lt;p&gt;By that test, seven products qualify right now: OpenAI Dots, Meta Muse, Grok Bot, Manus Cue, Gemini Spark, Claude Cowork and Perplexity Computer. Character.AI, Pi and Replika don't — they're excellent companions, but they wait for you. (Dot, the 2024 personal agent, shut down in October 2025 — a reminder of how fast this category moves.)&lt;/p&gt;

&lt;p&gt;The full section, with what each one actually does: &lt;a href="https://tools-hub.sweetai.work/best-ai-agents" rel="noopener noreferrer"&gt;https://tools-hub.sweetai.work/best-ai-agents&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Curious where you'd draw the line — is a browser agent that only acts while you watch an agent, or a macro with good PR?&lt;/p&gt;

</description>
      <category>ai</category>
      <category>agents</category>
      <category>productivity</category>
      <category>discuss</category>
    </item>
    <item>
      <title>I ranked AI models using only public benchmarks — and stopped publishing scores</title>
      <dc:creator>XM Yang</dc:creator>
      <pubDate>Tue, 27 Oct 2026 06:29:00 +0000</pubDate>
      <link>https://dev.to/xmyang0223/i-ranked-ai-models-using-only-public-benchmarks-and-stopped-publishing-scores-3hf5</link>
      <guid>https://dev.to/xmyang0223/i-ranked-ai-models-using-only-public-benchmarks-and-stopped-publishing-scores-3hf5</guid>
      <description>&lt;p&gt;Most "best AI" lists publish scores nobody can reproduce. Different agent harnesses, different prompts, vendor-reported numbers mixed with independent runs — the numbers look precise and mean almost nothing.&lt;/p&gt;

&lt;p&gt;So on Tools Hub's rankings pages I made two rules. First: order only, no scores. Second: every board names its source on the page — SWE-bench Verified (independent runs compiled by Epoch AI) for coding, HealthBench for medical questions, LMArena's blind human votes plus EQ-Bench for writing, and Arena's text-to-image votes for image models.&lt;/p&gt;

&lt;p&gt;The interesting part is where the boards disagree. EQ-Bench's judged board puts GPT first for creative writing; LMArena's human voters put Claude first. I follow the humans and say so on the page. That disagreement is more useful to a reader than a fake-precise 9.4/10.&lt;/p&gt;

&lt;p&gt;Four boards are live (coding, medicine, writing, image). Happy to hear what other public benchmarks are actually worth ranking from — that's the bottleneck, not the page template.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://tools-hub.sweetai.work/rankings" rel="noopener noreferrer"&gt;https://tools-hub.sweetai.work/rankings&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>machinelearning</category>
      <category>productivity</category>
      <category>discuss</category>
    </item>
  </channel>
</rss>
