Huge shout-out to ZCode's remote control — the gold standard of commanding a local agent from your phone.
The setup cost is nearly zero: register an account on your phone, no membership needed. Scan the QR code on your computer, then link it to WeChat or Feishu while you're at it.
The best part is the free built-in tunneling — you never have to think about your network environment. Lie on the couch or sip coffee out somewhere, send a message from your phone, and your computer's agent starts running code on demand.
Tags
Filter by theme
08.05.2026 · 7 min
DeepSeek V4 Flash stable release, hands-on: cheap, fast, 1M context, built-in search
PinnedA few days deep-diving the stable DeepSeek V4 Flash release: aggressively cheap, lightning fast, 1M context, official built-in web search, fully open source. The one weakness is multimodality — but you can patch it by combining other models.
DeepSeekModel reviewAI tools
09.09.2026 · 8 min
The First-Party Harness Isn't Always the Best Answer: In the Open-Model Era, Pick Your Model and Harness Separately
Using Claude? Go Claude Code. In the open-model era that instinct deserves an upgrade. The first-party advantage is real, but training a good model and building a good harness are two different kinds of engineering — here's why models and harnesses are now separable choices, plus a positioning map of five harnesses.
AI AgentHarnessTool selection
09.09.2026 · 8 min
Stable, Fast, Productive, Cheap: My AI Coding Stack, Documented
OpenCode + OpenChamber + two DeepSeek V4 Flash subscriptions: 15 projects running code simultaneously and the quota barely moved. The full configuration laid out — context pruning, layered memory, desktop automation, and the "stick with the original pairing" lesson.
OpenCodeAI CodingConfiguration
09.09.2026 · 7 min
TUI Is Killing Your Cognitive Bandwidth: Time to Shatter the "Geek Filter" in AI Coding
A pile of AI coding agents are rushing to ship TUIs, stuffing interaction back into the 1980s terminal paradigm and calling it "immersive" and "respectful of programmers." This piece breaks down the three mechanisms by which TUI systematically lowers your cognitive bandwidth — and why Web UI is the right answer.
AI CodingInteraction designEssay
Someone in a group chat asked today: is there a universal AI coding desktop? One that connects to every subscription and never dumbs down any model?
After talking with a few friends who tinker at high intensity every day, our conclusion was unanimous:
Stick with the original pairing.
Put a model inside a third-party shell and you basically can't escape degraded intelligence, slower responses, and broken subscriptions. Our setups all look almost the same now:
GPT → Codex: direct official subscription, the most stable native scheduling.
Gemini → AntiGravity: the speed ceiling, long windows without choking.
DeepSeek, GLM and other open models → all of them go to ZCode.
On why ZCode is a must-have:
ZCode makes it easy to configure third-party provider models.
ZCode's phone remote-control mode is really, really good — you get work done from the couch while scrolling your phone.
ZCode's cache utilization is extremely high.
People's takes on the Gemini "US Soybean" plan are packed with substance, so I put together a quick recap of the comment section:
Many people find Gemini disappointing because they mistake its identity — they treat it as their primary engineering machine tool.
In fact it's positioned more like a fully loaded Swiss Army knife: light, versatile, with rock-bottom marginal cost, but with clear limits on complex engineering and autonomous execution.
On the information-input side, native YouTube parsing is one of its most hassle-free abilities. No local download scripts or transcription tools — submit a link and it extracts the full timestamped transcript, keyframe grabs, and structured summaries in the cloud.
A lot of people's workflows have settled into a clear division of labor: Grok tracks real-time discussions on X, while Gemini handles tech podcasts and in-depth videos over half an hour long.
In visual and front-end scenarios, its multimodal foundation is solid. Full pages of handwritten notes, complex tables, or low-resolution scans all get recognized with high accuracy and few hallucinations.
And when prototyping front ends it's extremely fast, with natural code structure and none of that stiff, redundant generated feel.
Plenty of indie developers use fixed prompts to have it generate slide structures on a single page, or quickly produce usable pages from design screenshots.
On cost and tooling, its ecosystem math works out well: the $20/month subscription supports family sharing for up to 6 people, comes with terabytes of Google Drive storage, and integrates deeply with Docs and Sheets.
Route research, multimodal parsing, and light page building to it, then hand the core complex code to dedicated coding models, and your overall tooling costs drop noticeably.
A few advanced uses that have been proven to work:
First, leverage its relatively straightforward logical preferences and have it review the over-engineering and redundant code other models tend to produce.
Second, pair it with NotebookLM to build a private knowledge base around a specific set of materials.
Third, allocate the premium quota to derived needs like speech synthesis and image processing.
But you must be clear about its limits.
One, it's not suited for complex terminal-agent tasks; ask it to modify system rules or deploy environments and it often refuses or fails to follow multi-step instructions.
Two, deep business code must be cross-reviewed by your primary coding models to avoid hidden defects. Three, it's sensitive to your network environment — unstable nodes easily trigger service restrictions or a degraded experience.
US Soybean plan, activate!
Are these spam accounts mutating?
The garbage posts they're sending look different from before.
No problem — we update the list in real time and keep learning these fingerprints in real time.
It's the elite ones we're hunting!
After many rounds of iteration, FeedSieve — the X-feed cleaner — is officially live on the Chrome Web Store.Who wouldn't want the thrill of one-click blocking those spam accounts?Go for it, folks!
Curious what everyone uses the American soybean Gemini for?I feel like most people with Codex don't really need it day to day.What are its actual strengths?
With the FeedSieve plugin on, I like to go grind spam in the comment sections of hot posts — it's oddly therapeutic.
Just caught 33 spam accounts in one hot thread.
Blocked them all in one click!!!
An 18-month-old American soybean Gemini, bought for 12 yuan off Xianyu!Start it up!
This is my current setup.For Chinese models, I recommend mounting ZCode (DeepSeek V4 Flash, GLM 5.3 Flash, that kind).For GPT, go straight to Codex.Want speed? Use Pi.Want speed AND peace of mind? Use OMP.Building custom client-side apps for enterprises? Use DSH.Want to train your patience? Use Workbuddy.
Hard to imagine: someone locked up five years ago, released this year —looking at this AI world, what would they be thinking?
Yesterday I posted a comparison of Command GOAT and OpenCode Go, and Command Code's CEO personally came to comment.He said I'd missed a line item — the caching bill. I dug into it, and Command Code has something going on.A Coding Agent is essentially a giant Prefix Cache machine. A session of several hundred thousand tokens — by the next round, the genuinely new tokens might be only a few thousand. The system prompt, tools, code context and conversation history barely change. The higher the cache hit rate, the cheaper those old contexts get.Command Code recently did a serious round of Cache Thrashing work, mainly with these strategies:Pin sessions to the same GPU pod to keep the KV cache warm; keep the stable prefix as intact as possible and stop provider fallbacks from messing with context; plus their Tool Call Repair, where the harness fixes wrong model parameters locally instead of sending errors back to the model to retry.The v0.33.0 changelog literally lists: Major Cache Overhaul, median cost reduction 2.9×. They now officially claim sustained cache hit rates of 95–99%.Same model, same $10 plan — but because of caching, the difference compounds into a completely different outcome.
I took another careful look at Command Code, this time comparing it side by side with Pi, OMP, DeepSeek Harness, and OpenCode.
What makes Command Code valuable right now is the Harness Engineering it does for open models.
Pi's philosophy is minimalism: the harness interferes with the model as little as possible — light, fast, low token overhead, and extremely malleable. OMP keeps stacking heavy coding capabilities on top of Pi — LSP, Debugger, Browser, AST — closer to fitting the agent with a full IDE.
DeepSeek Harness goes even further: Everything is Plugin. Agent loop, tools, permissions, presets, UI — all can be taken apart and recombined, which makes it perfect for studying the next generation of agent runtimes.
OpenCode is the most balanced of the bunch: open source, many providers, a big ecosystem, mature Client/Server, desktop app, and subagents — a generalist that plugs into any model and handles any job.
Command Code takes a different route: it loves cleaning up after the model. Bad tool-call arguments get fixed locally by the harness; files read once and read again get deduplicated; long sessions keep a stable prefix to raise cache hits; when context is about to blow up it runs compaction; plus session affinity, taste, subagent isolation, and so on.
Applied to workhorse models like DeepSeek V4 Flash or GLM-5.3 Flash, this approach gets extra value: the weaker the model, the more the harness picks up the slack.
So here's how I'd position all of them now:
Pi — best as the foundation for your own long-term agent: simple, clean, endlessly hackable.
OMP — best for genuinely heavy coding: LSP, debugger, complex repo navigation.
DeepSeek Harness — best for tinkering with agent architecture, presets, multi-agent setups, and next-gen runtimes.
OpenCode — best for anyone who wants a mature, general-purpose, multi-model coding agent.
Command Code — best for heavy users of open models like DeepSeek / GLM, especially anyone running Flash workhorse models at scale.
If I only care about long-term malleability, I still prefer Pi. But if today you asked me to put DeepSeek V4 Flash / GLM-5.3 to serious work, I'd genuinely give Command Code a taste.
I got so annoyed by garbage accounts like "Fu Bu Hei" lately.In a fit of anger, I wrote a browser extension — it's called "FeedSieve".After several rounds of iteration it's in a good place, and it's about to go live.For those who need it, bookmark this — I'll open-source it and put it on the Chrome Web Store soon.
For folks who regularly use Chinese models like DeepSeek and GLM, it's worth checking out the "family bundles" foreigners have been putting together:Command GOAT and OpenCode Go.I went through both $10 plans carefully.The conclusion: pick Command GOAT with your eyes closed, no question. As of late August 2026, the version answer is:The Command GOAT $10 plan!!!
https://commandcode.ai/docs/plans/goat
The "amazing $5" OpenCode Go everyone was hyping on X has completely fallen apart.OpenCode Go used to have a first-month $5 deal that kept its value-for-money alive — now even that is gone, $10 from the start.By comparison, the competitor's Command GOAT plans are starting to look better and better.DeepSeek V4 Flash, GLM-5.3 Flash, GLM-5.3 are the workhorse models for heavy coding and running agents right now — and the Command GOAT plan gives them noticeably more generous quotas.I now split coding models into two simple buckets: the core models and the grunt models.V4 Pro, GLM-5.3 and the like are core models — they chew the hard bones. Complex architecture, gnarly bugs, big refactors, key decisions: when a problem is truly hard, hand it to them.V4 Flash and GLM-5.3 Flash are grunt models. Reading repos, searching code, editing files, running tests, batch work, breaking new ground — that dirty work is exactly what Flash should grind through.The Command GOAT plan has plenty of grunt quota, and its core model quota has no obvious weak spot either.So I'm declaring it: the Command GOAT plan is the current version answer.
Get a $200 ChatGPT Pro membership, use it well, use it clean.You're in AI.
DeepSeek Harness + third-party API compatibility issue:When the agent calls tools like Bash, it errors with
ToolNotFoundError: unknown tool "", which looks like the tool was never registered.What actually happens: the gateway returns a different data format from DeepSeek's official API, and the tool name gets dropped somewhere in between.A few relevant GitHub Discussions links:- Search "unknown tool" in deepseek-ai/dsh#issues
- People running self-hosted gateways (OpenRouter / OneAPI / new-api) are the most likely to hit this
I went back through the DeepSeek Harness repo and Discussions recently, and this project is at a very interesting stage:DeepSeek Harness already has 150k+ stars and over 12,000 commits in the repo.Yet the official README still states very clearly: it's only a Developer Preview right now, and breaking changes are still coming.rc.7 was released on August 17. Issue and PR creation is restricted; feedback goes through Discussions.The community is huge, but core development is concentrated in few hands. Bugs keep popping up — like the unknown tool "" issue.What's most missing: test coverage, provider compatibility, and session reliability.Architecturally it's still attractive, like the foundation of an agent operating system:
- Everything is a Plugin
- Cordis composition architecture
- Preset / Skill / MCP / Agent / Provider all pluginized
I'm increasingly convinced that the value of emotion is deeply valuable.Why can we Homo sapiens band together to do great things?Precisely because we collectively believe in some story, some emotion, some totem, some myth, some narrative — that's what forms organizations, drives cooperation, grows us stronger, and makes greatness happen.Isn't that itself a kind of emotional value?
Lately I've been studying DeepSeek Harness. There's a concept called Preset worth paying attention to.Preset can be understood as the distribution unit of an agent. Load different Presets on the same model, and it becomes completely different agents.What's the biggest difference between a Preset and a Skill?
- A financial-report-analysis Skill: teaches the agent how to analyze financial reports.
- A financial-report-analysis Preset: bundles the financial-report-analysis Skill, web search, Browser, PDF, Python, subagents, research prompts, and context strategy all together — handing you an out-of-the-box agent.
- One: a Plugin / Skill marketplace, mainly for developers and power users building agents;
- One: a Preset / Agent marketplace, truly aimed at ordinary users.
DeepSeek Harness's ecosystem will very likely boom through Presets.A Preset plaza has already appeared in the community — interested folks can keep an eye on it.Entry: dshdesktop.com/preset
What's scarce is a feel for building products, an insight into demand, a mind that understands users.
This article makes the case: when using DeepSeek Harness, try PTC mode as much as you can.
The benefits are faster speed, lower token spend, and a much more focused model attention.
I spent some time on DeepSeek Harness's four modes — here's the pros, cons and use cases for each.
- Standard — best for everyday development Standard is DeepSeek Harness's default full Coding Agent. It ships with everything everyday development needs, and it's especially suited to tasks where you don't yet know where the problem is and the agent needs to explore on its own.
- PTC — the efficiency mode for complex tasks PTC lets the model write a block of TypeScript that combines multiple tool calls into one execution.
- Minimal — just Bash + Editor for the model Its biggest strength is simplicity: few tools, small tool schemas, low fixed prompt overhead.
- Creator — using an agent to create agents It has Standard plus the Cordis Runtime, plugin experiments, and Preset creation. It even turns the agent itself into something composable and creatable.
Awesome DeepSeek Harness Desktop has arrived.
ADHD — an outstanding DeepSeek Harness desktop build, an open-box Electron wrapper made for DeepSeek Harness. Awesome DeepSeek Harmony Desktop has arrived.
Original post: https://x.com/zacklin233862/status/2087922286368375265?s=20
A pile of AI coding agents are rushing to ship TUIs these days, and Claude Code owes a big share of the blame for that trend.
It made conversing with an AI in the terminal to write code an art form.
The problem is, somehow a "truly advanced" coding agent now has to default to a terminal interface — the web version becomes an afterthought, even mocked as not programmer enough.
And the result?
Build a complicated TUI first:
all the colors, status bars, shortcuts, and mode switching,
then call it "immersive," "efficient," and "respectful of programmers."
Really?
For the small group who live in the terminal all day, sure, it's great. But for many more people it's a wall built on purpose: features hidden behind keys and modes, new users completely lost, and even veterans relearning every tool's personal keymap.
A web UI can be just as fast and clean — install it as a PWA and it's nearly indistinguishable from a desktop client, with easier updates too. Forcing interaction back into the 1980s terminal paradigm and then sneering at the web is mostly self-indulgence.
Claude Code itself is excellent, no dispute there. But it popularized the taste and the path dependence of "TUI is the orthodox way," and the imitators that followed are even worse.
Tools exist to serve people, not to prove who is more "hardcore."
Dressing up anti-human interaction as sophistication — it's genuinely hilarious.
Honestly, some geek communities aren't just niche anymore — they're outright solipsistic.Their signature move: treating inconvenience as sophistication, and barriers to entry as taste.If something is easy, intuitive, and quick for normal people to pick up, they instinctively feel it's not hardcore enough, too mainstream, beneath them.Conversely, the more awkward the interaction, the more buried the information, the more you need to learn a private jargon and a pile of shortcuts — the more they feel this is what a real tool looks like.This isn't about efficiency anymore. It's using pain to confirm your own identity."I can endure interactions other people can't" — once that sentence leaves your mouth, the superiority writes itself.The smaller the circle, the easier this self-congratulation feeds on itself.It eventually turns into: whoever dares to call it unusable simply doesn't get it — they're an outsider.Truly capable people make complex things simple. Some geeks make simple things complicated, then stand on top of the complexity looking down at everyone else.
I think DeepSeek Harness nailed the form factor.Instead of jumping straight into an anti-human TUI, it ships a Web UI.And it barely takes any storage — because there is no installer at all. No Electron framework ramming a browser engine inside. That's a lot of storage saved.The only drawback is that it lives in the browser — but you can install it as a PWA, and then it feels exactly like a native client.Installing a web app as an application via Chrome is genuinely useful. In the screenshot here, it's a webpage, yet it behaves no differently from a desktop app.
So well put — this is the GPT-5 generation's endemic flaw:It keeps writing internal requirements into the UI copy to prove its workload. To prove itself — oh look, I really did it, aren't I great, I didn't slack off, I've laid every one of your requirements out on the table.And the UI copy is bloated beyond belief. From 5.0 all the way to 5.6 — generation after generation — this still hasn't been fixed. It's in the marrow.Anyone who's actually shipped frontend products knows:
- Every bit of text on the interface, big or small, is user-facing copy
- Text that belongs in chat feedback to developers should never be written into the product
The reading site for the English podcast is finally out.Its name: Cherina PodCome visit. Give me your harshest feedback — I'm going to iterate hard, fix hard, optimize hard.
https://pod.cherina.app
OpenChamber shipped another new version. You can feel how much the author cares — this may be the most lovingly maintained front end I've seen in an open-source project.
The OpenChamber + OpenCode + DeepSeek V4 Flash combo I run every day honestly feels fantastic.
Every time I see a Pi Agent related project, I get excited.
The workflow built with herdr + pi deserves a proper deep dive.
Qwen shipped an official plugin that gives DeepSeek V4 Flash a pair of eyes — try it if you're interested.
I've tried pretty much every open-source Harness GUI client on the market, and most of them are hard to look at, let alone enjoy.
OpenChamber is the best-polished open-source Harness GUI client I've used: refined, stable, and the occasional bug doesn't get in the way of long-term use.
But it only works with OpenCode and can't switch to other harnesses. Some users report that OpenCode's overall performance isn't great and can hold the model back.
So I started sketching a project: keep OpenChamber's front end as the skeleton and swap the core for Oh My Pi, which better fits my own needs.
Why not the native Pi Agent, and why Oh My Pi instead?
Because the native Pi Agent is too flexible — there's no way to build a general-purpose GUI around it.
As a compromise, Oh My Pi works like a Pi Agent distribution: the plugin interaction experience is relatively stable.
So I think turning OpenChamber + Oh My Pi into an out-of-the-box product is a real, unmet need.
I named the project OMPChamber. It's still in intensive development and testing — star it if you're interested:
https://github.com/realchendahuang/OMPChamber
We really should pay more attention to designers and programmers with taste.
DeepSeek V4 Flash is genuinely smart, and genuinely cheap.
I've had 15 projects open in a row, all running code at the same time.
The quota barely moved.
Sharing the OpenCode setup I'm currently running. It's built around three core needs:
Context management — long sessions that don't blow up the context, saving money and tokens
Memory — remembering user preferences, project decisions, and past lessons across sessions
External capabilities — docs lookup, code search, desktop automation
The main plugins I use:
- opencode-dcp (open source) — dynamic context pruning plugin. When the context hits a threshold it automatically compresses older content into technical summaries, keeps the key information, drops the noise, cleans up duplicates, and removes low-value content like tool errors from the context.
- opencode-goal-plugin A Goal-mode plugin that adds goal management to long tasks.
- Hermes Memory (I'll open source it if there's demand) My heaviest plugin — I ported the Hermes agent's layered memory mechanism onto OpenCode.
- context7-MCP Look up the latest official docs for libraries and frameworks.
- grep-MCP GitHub code search — find real-world usage across all of GitHub.
- open-computer-use (desktop automation) Lets the AI directly operate macOS apps — clicking, typing, scrolling, dragging, screenshots, reading the accessibility tree.
08.06.2026 · 6 min
After tinkering with Pi Agent, OMP, Codex, and ZCode, why I finally chose OpenCode + OpenChamber
A retrospective on choosing an Agent Harness: three criteria — GUI experience, vendor lock-in, and freedom for secondary development — ruled out Pi Agent, OMP, Codex, and ZCode, and finally settled on the OpenCode core + OpenChamber interface.
AI AgentTool selection
08.06.2026 · 5 min
Giving DeepSeek multimodality: the Qwen-3.7-Flash vision approach
DeepSeek V4 Flash has no multimodality — what do you do for image understanding? After surveying the market, the most cost-effective option right now is Qwen-3.7-Flash: recognizing an image costs almost nothing, and pairing it with V4 Flash patches the gap.
DeepSeekMultimodalityModel review
08.05.2026 · 6 min
DeepSeek's API has built-in web search — freeload the official search via the Responses API
DeepSeek built web search straight into the API: call deepseek-v4-flash through the Responses interface and declare the web_search tool. No third-party search engine integration, no search API key required.
DeepSeekAPIAI tools
The OpenCode Go plan is just $5 for the first month — 35 yuan.
Once you're in, use its new DeepSeek V4 Flash model.
In practice: I pounded it all morning with heavy use and only burned 1% of the weekly quota. What the hell — I can't even use it all up, now I'll start getting anxious again hahahaha.
The impossible triangle of cheap, good, and fast has been perfectly achieved by the OpenCode Go + DeepSeek V4 Flash combo.
Opencode https://t.co/c0k1WKWNle
I'm genuinely in awe — OpenAI has lost its mind. GPT-5.6-Luna price cut 80%: $0.2 input, $1.2 output, $0.02 cache hit.
Compared with DeepSeek V4 Pro, at typical coding usage — 100M tokens with 95% cache hits — DeepSeek runs about 20 yuan, Luna about 26 yuan.
And if you're on a Codex plan, running Luna gets you five times the previous volume — a full fivefold usage boost! https://t.co/fBVVfyN2OC
Liang Wenfeng said at an investor meeting:
"The endpoint of intelligence may all be embodied. Because for a normal person, what they need isn't a computer — it's labor."
Those words are spot on. Ordinary people have never wanted a smarter computer; they want "labor" that actually gets work done. No matter how capable text is, it's only an intermediate state. https://t.co/2xg5o5hNjt
Let me introduce a concept: FOMO, short for Fear of Missing Out.I'm sure plenty of people have this problem: you see something good and want to bookmark it, you see a new tool or trend and want to jump on it, always feeling that if you don't, you'll miss something important. The result: your bookmarks get fuller and fuller, your head gets messier and messier, and your attention gets nibbled away bit by bit along the way.
People configuring AI tools usually set an API key, and sometimes you're asked to pick an interface format.The interface format is, simply put:the agreed way of communicating between you and the model service.For example, the "OpenAI-compatible interface" most people know often refers to Chat Completions. That format excels at chat and simple text generation, which is why model vendors everywhere prioritize compatibility with it.But building AI apps today is far more complex. Models now need to read files, browse the web, call functions, run code, connect to MCP, hold context, and keep going on the next steps.That's where OpenAI's Responses API earns its keep.
- Its reasoning models perform better. The same model, placed in a more fitting interface structure, completes tasks better.
- It's natively suited to agents. One API request can call multiple tools.
- It costs less. Compared with Chat Completions, Responses API cache utilization improves by 40%–80%. Higher cache utilization means cheaper repeated calls and easier latency optimization.
Just thought of a rather interesting angle.Zodiac signs, birth animals, astrology, fortune-telling — these can perhaps be understood as "compressed models" from humanity's low-compute era.The ancients had no databases, no statistical tools, and no AI.All they could do was observe seasons, celestial phenomena, farming calendars, personality, social conventions, and societal cycles over long stretches of time, then compress that complex experience into a symbol system easy to remember and easy to spread.The 12 zodiac signs, the 12 birth animals, the Eight Characters, the hexagrams — they're all essentially ways of reducing a complex world to lower dimensions.Of course, they can't be taken as rigorous scientific prediction.But they're not necessarily pure superstition either.They're more like a distillation of experience — built by the ancients, with stories, symbols, and rules, when compute was scarce, data was scarce, and tools were scarce.
My X-scrolling frequency has dropped lately.Not sure if I've entered sage mode.After some honest reflection, I realized I care about one question more than anything:can the stuff I scroll past every day end up as a real product?Watching a hundred tool updates is no match for breaking through one concrete scenario.These days I want to put more energy into products, community, and long-term content.
Lately I've been rethinking knowledge management.I realized there are actually two paths here.One is building on Feishu.Its advantages are direct:free, multi-device sync, shareable, easy collaboration, and rich enough in form.Documents, multi-dimensional tables, spreadsheets, slides, canvas, Todo, knowledge bases — these modules combined already cover a lot of personal and team knowledge-management scenarios.I'm especially bullish on the multi-dimensional tables.They're well suited as a knowledge hub.Information collection, material archiving, status management, structured organization, content re-production — all of it can revolve around them.Add the Feishu CLI and AI agents, and Feishu stops being just an office tool.It becomes more like a knowledge workflow system agents can operate.The other path is building your own on Cloudflare.This path is freer and better suited to productization.FlareMo, the thing I'm building now, is an attempt at a knowledge-management product on Cloudflare.Cloudflare's advantage is that the underlying capabilities are complete:frontend, backend, database, object storage, vector search, AI — all combinable.But it takes development.And you have to design the data structure, search experience, permission system, and product logic yourself.So my current take:
- Feishu is for quickly standing up a knowledge-management system you can use today.
- Cloudflare is for turning knowledge management into a truly controllable, iterable product.
What is knowledge management?Mention knowledge management and most people's first reaction is: bookmarking articles, organizing links, taking notes, building a beautiful library.But from first principles, the problem knowledge management actually has to solve is:Can the information I see today keep helping me make judgments in the future?Human memory is limited, attention is limited, and experience slips away easily.If a piece of material merely gets bookmarked, it'll likely end up as an information graveyard.Truly valuable knowledge management should complete one transformation:turn external information and personal experience into an action capability that's retrievable, callable, reusable, and updatable.
- An article, merely bookmarked, is just material.
- You distill its viewpoint and it becomes information.
- You connect it to your own projects, scenarios, and judgments — only then does it start becoming knowledge.
- Finally, when it helps you write articles, build products, make decisions, and design workflows, it becomes capability.
Remote into my home Mac via NetEase UU Remote.Code anytime, anywhere.Keep my wife company shopping, anytime, anywhere.
Sure, the most eye-catching thing is an indie developer's income.But for us ordinary people, the most valuable thing is still the path they walked to get there.
What black magic does the Brave browser use?Zero ads on YouTube, like magic!!
A friend asked me what brand the power bank in the photo is.Alright, time for me to start pitching.This is the CoolState Power Bank No. 25 Super SE (CoolState is the former Zimi).A little over 140 yuan, 25,000 mAh, up to 120W output, built-in charging cable.It keeps my MacBook Air going through a whole afternoon of heavy work.The perfect office companion for Starbucks, McDonald's, KFC, the countryside, and every other place!
Here's a fun fact — you might assume DeepSeek is selling tokens at a loss.Its API is priced cheap, but the official side has mentioned a profit margin of 545% on this segment!!!So DeepSeek is not only a cyber-AI bodhisattva, it's also technically a monster.Original: DeepSeek Open Infra Index
Holy shit! I only just found out the Firefox browser gives every user 50G of free VPN traffic per month.Extremely handy for that awkward chicken-and-egg situation.How to use:
- Type
about:configin the address bar - Go to advanced preferences and click "Show all"
- Search for
browser.ipProtection.enabled - Flip that entry's value to true
Gave Tencent's Workbuddy a spin — there's this classic scenario: claiming Meituan coupons.At the login step, it generated a QR code for me to scan and sign in.Then I just barked "go claim my coupons" and it went and claimed that many for me.Using an AI agent for something like claiming coupons — absolutely amazing.
Sharing a treasure of a site: BuzzingIt's for anyone who wants to widen their information sources and bust out of their echo chamber.It continuously pulls hot content from Hacker News, Reddit, BBC, Reuters, Bloomberg, Product Hunt and other foreign sources, and translates the headlines into Chinese.Its biggest value: you get to skim the headlines in Chinese first.If something catches your eye and looks worth reading, click in for the details.At that point, pair it with Immersive Translate for bilingual reading.
If you've bought a circumvention airport and also built your own VPS nodes, and you want them managed uniformly, routed uniformly, and output as one subscription link.
You want routing rules, node filtering, and combination logic maintained on the server side instead of configured client by client.
You want to run it with Cloudflare's lightweight deployment — no servers, no database services, no complex background jobs to maintain.
Sharing the browser extensions I've kept close for years.
- Wetab: a new-tab extension. Put your frequently used stuff in it, customize search engines, and sync across devices free with an account. I built a Google-based X search engine in it — way better than the built-in one.
- Video Speed Controller: fine-tune video speed, great for absorbing knowledge efficiently.
- XHunt: a must-have for X. See the leaderboard for each KOL and the hottest posts right now.
- X Mutual-Follow Radar: marks mutual follows and one-way follows on the page, and can show the following / followers ratio.
- FireShot: capture webpage screenshots, edit, and save as PDF, JPEG, GIF, PNG, or BMP.
- Proton Pass: an elegantly designed password manager with multi-device sync.
- Immersive Translate: the veteran bilingual translation extension, wired up to the DeepSeek V4 Flash API.
- Codex: essential for connecting Codex to Chrome. With it you can automate anything in the browser — like writing an appeal email when your account gets banned.
06.28.2026 · 4 min
I moved my proxy subscription aggregation to Cloudflare
OriginalMerge multiple proxy services plus your own nodes into one single subscription, with routing rules configured server-side — clients just subscribe.
CloudflareCircumventionOpen-source project
Humans do the judging. AI does the amplifying.Human attention is limited. Humans handle only the most critical slice of judgment.You have to spend your limited attention where it's worth something, and let AI take on more of the grunt work.That's the whole point of AI workflows.Let the AI workflow swallow all the grunt work; humans handle final judgment — the review or acceptance of the most valuable, most core, end-goal results.That's the state I want to reach. I've been casually researching these workflows lately too.
A notes app natively built on the Cloudflare tech stack — FlareMo.Register a free Cloudflare account, and once deployed you get 10G of file storage and 5G of data storage.Free forever — no cost for infrastructure, no cost for traffic, data backed up across regions, high availability built in from birth.Serving in the cloud 24/7, fully compatible with the Memos ecosystem.Development progress at the time: got the core functionality up in a day, already fully usable.Going forward: more usage and deployment tutorials, engineering test cases, AI-agent techniques, and multi-platform support. Build in public.
I moved my circumvention subscription aggregation to Cloudflare.Open-sourced a subscription-link aggregation project that deploys completely free on Cloudflare.It folds all your circumvention subscription links and self-built nodes into one link, with routing rules attached.Truly one link — usable in every client, no per-client configuration.GitHub: https://github.com/realchendahuang/sub-store-cloudflareFull post: I moved my airport subscription aggregation to Cloudflare
Clash Party for circumvention
- Substore deployed on Cloudflare one subscription link handles multi-provider aggregation and rule-based routing feels so good (that second tone)! Also plugging Clash Party as a client — it's very elegant. My sub-store-cloudflare open-source repo: https://t.co/u8zhrOSrAw https://t.co/Zzc3IXKsLI
Took me a while, but I got Sub-Store deployed on Cloudflare.Merged my self-built nodes and the airport's backup nodes into one, and paired it with ACL4SSR routing rules on auto-subscribe.Done with a single subscription link:
- Merging nodes from multiple providers
- Node filtering
- Routing rule configuration
- Automatic config updates
Lately I've been thinking about a plan.
Moving my personal knowledge management, information collection, and knowledge output entirely onto Cloudflare.
A free account already comes with 10G of object storage and 5G of database.That's basically a self-hosted, cloud-based, permanently free knowledge management system.
Let me tell you about my company-naming saga.I used GPT 5.5 and Gemini 3.1 Pro, brainstorming with both models over and over — two or three days, and still no good name.Why? Because these models have no creativity. Every name option they gave me was just piles of fancy words and character combinations.You throw out your own idea, and they roll the same circular talk back at you, round and round, until my blood pressure spiked.There's another big problem: once you finally land on a few names you're happy with, you open Qichacha and — damn — they're either registered already or eerily similar to something existing.Here I can recommend Alibaba's 88 Cha, like a free Qichacha — use it to check for name clashes.If you don't want to keep repeating existing names, go with a four-character name, and keep your thinking broad.Above all, switch between several models — several different brains.Feed the AI some names you consider great as references, like DeepSeek (深度求索) or ByteDance (字节跳动).Let the AI study the names you think are good. Only by borrowing from the best of many can a truly great name emerge.
Looking for a tool that lets you "jot things down casually and slowly grow them into knowledge."I want Flomo's out-of-the-box, zero-friction inspiration capture;I want Memos' mature open-source ecosystem, portable data, and an API you can actually play with;But I don't want to spin up another VPS and babysit Docker, Postgres, and a Node process that runs forever.So what to do? FlareMo's approach: treat Memos' data model and API as the "spec," and rebuild the whole runtime on Cloudflare:Workers run the API, D1 stores notes, R2 holds attachments, Queues/Cron handle background jobs, and Vectorize + Workers AI are saved for future semantic search and AI workflows.Standing on Memos' shoulders to build a personal knowledge base that lives on the edge cloud.
A Cloudflare Container victim has appeared.Because the code I was writing was still in the debugging stage, the stop command never took effect.So one Container ran continuously for 7 days and cost me $4.
The name for the Eastern-Western mythology compendium is set — God-Museum · Museum of the Gods:Cram the Classic of Mountains and Seas, Journey to the West, ancient Greek, Norse, and ancient Egyptian pantheons — these sets of stories that never touched — into one universe, and make them not collide.The same law of thunder is called Zeus in Greece and Leizhenzi in the East;the same great flood is remembered in the East as Nüwa mending the sky, and by the Hebrews as Noah's Ark.If you keep chasing that thread seriously, the whole mythological world has to be rearranged.Try it online: https://chendahuang.com/god-museum/
Where do the new software opportunities of the AI era lie?A big one: the part that traditional software can't record — in other words, unstructured background data.For example, a truly AI-native CRM, when logging a discount given to a client, can also log why this discount was given instead of another, why the other options were rejected, and who talked whom into it.Traditional software has a huge amount of hidden information like this that it simply can't record. The AI era, with its ability to understand text and video, can pull this hidden, highly redundant information out into structured data for analysis.
Day one moving into the OPC community
From now on, this workstation is mine
Nice! https://t.co/3rSQcZ760R
Today's my first day officially moving into the OPC office. Let me walk through how I started an OPC.My original motive was simple: have a compliant entity to receive payments and publish products. So I went online looking for company registration info.I wasn't picky about the city: Nanjing, Suzhou, Hangzhou, Ningbo — any developed coastal city was a candidate.By coincidence, I was in Nanjing for work one day, and while idly scrolling through Nanjing-related WeChat accounts, I found the phone number of an OPC community contact. So I added them.After a quick introduction of my project, the person in charge said OK and arranged a site visit the same afternoon.Their ask: work on-site three days a week. If I could accept that, they'd let me register the company here.Since this is under Nanjing's state-owned assets, the policy terms are outstanding — water, electricity, internet, and property fees all free, theoretically for three years. They even assigned a bookkeeping accountant to register the company for me, carve my company seal, get the business license, and set up the bank account. So touching it's almost absurd.After two or three weeks of the registration process, I finally got all the paperwork.The park manager told me to pick out a workstation. My bar for a workstation is low — just a seat, a stool, and power.Then they had me fill out three project application forms, mainly covering my project's strengths, the team's profile, and the AI-related edge. If the applications are solid and credible, there's a shot at a 50,000–100,000 yuan policy subsidy.Nanjing has genuinely handled this well. Everyone should check whether their own city has similar policies, or whether their provincial capital offers anything better.Good policies don't wait around — you have to go ask. By the time I arrived today, this park already had plenty of companies moved in; I'm actually on the late side.That's all for today. I'll keep sharing my startup journey and the retrospectives along the way.Thanks for the care, everyone.
Stating an opinion:Code is the leverage of us ordinary people.Codex is the leverage of leverage.Pull Codex out, and it's like pulling the firewood out from under the pot — the whole thing collapses.As they say: when the foundation isn't solid, the mountains shake and the ground trembles.Come on, domestic AIs — get your act together!
Lots of people use AI to write Cloudflare projects, and when it blows up, it's not about syntax.
The real problems:
AI writing Workers like they're Node.js
treating Bindings like process.env
forgetting await
stashing request state in globals
not knowing static asset requests are free
and calling its own R2 from inside a Worker over the REST API
I've collected all these pitfalls in the Cloudflare Playbook.
And by the way, I also worked on
For indie developers with no idea what product to build,start by satisfying your own needs.Chances are there are users on the same wavelength as you.
I made a somewhat ridiculous Skill:Dahuang AI ToneEveryone talks about "removing the AI flavor." But what exactly is the AI flavor?It does two things:
- Rewrite a piece of normal text into the AI flavor of a specified model
- Or the reverse: point out where a piece of text smells like AI, and which model's tone it resembles
Sharing over a year of hard-won, blood-and-tears indie dev pitfalls.In 2026, if an indie developer wants to build an app, the best practice is:Do Web + PWA first at the MVP stage. Fast to develop, fast to debug, fast to iterate.Web stack:React + TanStack Start + Vite + TypeScript + shadcn/ui + Base UI + Tailwind CSS + better-auth + Vercel AI SDK + oxfmt / oxlintUnified backend:Hono + Cloudflare WorkersWeb debugging:Chrome DevTools MCP + Codex Chrome extensionOnce the product genuinely runs, then do mobile.Mobile stack:React Native + Expo + Expo Router + NativeWind + React Native Reusables + Reanimated + Gesture Handler + EAS Build / Submit / UpdateThe core idea is simple:Validate the demand with Web + PWA first.Once you really have users, bring out React Native + Expo for the real app.
I spent a whole day properly researching "removing the AI flavor from articles."Compared the mainstream Humanizers, the de-AI-flavor Skills, and the prompt-based approaches out there.Then factored in my own experience from using LLMs since 2023 — tormented day after day by the AI flavors of GPT, Claude, Gemini, and Doubao.And made this:Dahuang Human ToneA de-AI-flavor Skill built for the Chinese context.My wish:No AI essay should be a pain to read.GitHub: https://github.com/realchendahuang/dahuang-human-tone
I made a Cloudflare Playbook.For every indie developer who wants to build products with AI coding.It covers:
- How to pick among Cloudflare's common services
- How to use Workers / Pages / D1 / R2 / KV / AI Gateway
- Whether the free tier is actually enough
- When to buy the $5 Workers Paid
- The pitfalls where AI-written CF code blows up most often
- How to set up Cloudflare Skills and MCP for Codex / Claude Code
- Which open-source projects you can reference directly
Broke 2000 followers on day 7 of operations — a quick look back.The core is one word: altruism.I mainly shared two kinds of practical content:
- A macOS heavy-usage software list (48k reads)
- A Cloudflare free-tier table + Playbook (decent numbers on X; some guy even scraped and rewrote it onto Zhihu, pulling several hundred bookmarks)
- Plus OPC one-person-company resource recommendations (38k reads)
I made a macOS Playbook.For people who just bought a Mac, just switched from Windows, or have always felt their Mac never quite clicked.It covers:
- What to install first after getting a Mac
- How Windows users adapt to macOS
- How to choose tools like Rectangle / AltTab / Homebrew
- How to work across devices
- How to set up the terminal, Git, SSH, and AI coding environment
- How to handle system maintenance, troubleshooting, secure backup, and device migration
06.20.2026 · 10 min
The founding-intention principle: why indie developers must learn to let go
Why do you set out? Who do you set out for? Which problem are you actually trying to solve? A product's sharpness comes from what you dare to cut.
Product thinkingIndie development
06.20.2026 · 8 min
Don't get carried away when users say "that's awesome": pain points vs. itch points
Verbal approval is worth nothing. There's only one hard standard for judging demand: what the user is willing to pay.
Product thinkingDemand judgment
A quick tech-selection breakdown of Docker options on Mac.
- Colima: free and open source, very lightweight, but basically command-line only. For people who don't mind tinkering and live in the terminal.
- Apple Container: Apple's new container direction — worth watching, but the ecosystem is still at a very early stage. Even Docker Compose isn't supported yet.
- OrbStack: has a GUI, light and fast, with solid ecosystem support. Downside: not open source — the commercial version needs a license, but the personal one is free.
- Docker Desktop: the official product, yet heavy and bloated, with a lousy experience. On a Mac, avoid it if you can.
I'm declaring it:Cloudflare is a true AI-native company.It's genuinely rebuilding the developer experience around AI agents.The just-released Temporary Accounts let an AI agent write code, deploy directly, get a temporary URL, test itself, and iterate itself.Smoother than smooth.Pair it with Cloudflare's official Skills, one command to install:
npx skills add cloudflare/skillsWhat does AI-native even mean?THIS is the GOAT of AI-native.A lot of friends are interested in OPC (one-person companies).While digging through research materials I stumbled on some treasure sites — sharing them with everyone:https://www.opcquan.com/communities
After years of deep macOS use, here are the pieces of software I've found to be genuinely great:
- QuickRecorder: a screen-recording gem — tiny, yet powerful.
- ClashParty: circumvention with the Mihomo core — lots of protocols, fast updates, and a good-looking interface.
- V2rayN: circumvention with lots of protocols; the UI is a bit dated, but it's stable and free.
- Doubao keyboard: voice input accurate to a scary degree. Supports macOS, iOS, Android; Windows is on the way.
- Brave: a privacy-first browser with built-in ad filtering — cleaner than extension-based filtering, and the best companion for watching YouTube.
- Mos: makes your mouse wheel scroll silky smooth.
- Amphetamine: keeps your laptop awake with the lid closed — the best companion for vibe coding.
- UU Remote: NetEase's remote desktop tool, free and fast. With it, what caps the image quality is never the membership — it's your network speed. I usually use it to remote into my Mac from my phone to check vibe-coding results; the key mapping is absurdly efficient.
- Tencent Lemon: Tencent's Mac junk cleaner — uninstalling apps is smooth, and the menu-bar icon is handy.
- Mole: the strongest Mac junk cleaner out there — terminal version free, UI version paid.
- Tailscale: WireGuard-based remote networking. With it, you can reach your little home computer from the ends of the earth.
- Obsidian: in the AI era, the local-first knowledge management tool. The ecosystem is just too good — pair it with the official Web Clipper browser extension and it works wonders.
How much of Cloudflare can you actually freeload on a free account?
And what pitfalls and best practices are there when using Cloudflare?
I spent a full three days digging hard into these questions.
The answers are systematically organized in the repo below.
I'll keep updating this repo:
https://t.co/gDuCwIdWHihttps://t.co/WZqEZLH9pE
06.18.2026 · 6 min
Debating whether something has value is itself worthless
Our value is not defined by anyone. In the age of AI, how should we understand ourselves?
AI thinkingHumanistic essay
AI has no memory — do we need to be polite to it?I read an arxiv paper claiming that if you curse at AI, it works harder.When we talk to AI, if we often say:
- "You're such an idiot. Don't slack off."
- "You piece of garbage, work properly."
- "I pay for a membership to feed you, and this is the crap answer you give me?"
- "Hurry up — if you won't do it, there are plenty of AIs that will."
- "Thanks, man"
- "Hey, buddy"
- "Good job, sweetheart"
A guy asked:is there a good write-up of Claude Code to study?My answer: in the AI era, always think with AI — let AI help you learn.If you genuinely want to learn how agents work,go look at the Pi Agent project.Small but complete — perfect for learning, and the code quality is high.Claude Code isn't open source; that's a leaked source dump.Pi Agent is the genuinely open-source one.
Today a friend shared some points in my comment section about equity and liability issues with one-person companies.First, many thanks to this friend for the kind heads-up.His view: you should have a family member hold at least 1% of the shares.The problem with a one-person limited company: when there's only one shareholder and you can't clearly separate company money from personal money, the shareholder may bear joint liability for the company's debts. Article 23 of the Company Law is exactly this logic.If the company has multiple shareholders, it shifts from a "one-person company" to an "ordinary limited liability company," dodging the special burden-of-proof risk of a one-person company and giving yourself an extra layer of low-cost protection.Hope this serves as a reference for everyone.But day to day, we should keep business and personal accounts clearly separate and aim to explain every significant sum.
- Where the money came from, why, and which account it went into.
- Where the money went, why it was spent, and who approved it.
- When you take money out of the company, is it salary, reimbursement, a loan, or dividends?
In the AI era, the most widespread form of procrastination:
hoarding tools, buying courses, joining communities, switching models, tweaking prompts.
Busy every single day.
Yet never daring to face one question:
do I actually have a real business on my hands that's worth throwing AI at?
Sharpening the knife keeps you very occupied.
But plenty of people never dare to actually kill the pig.
Sharing a small rewrite trick.Use the Doubao keyboard's voice-to-text.Keep the draft you need to organize in front of you.Speak it out, retelling it in your own words.That way there's less AI flavor.It's efficient, free, and needs no membership.You can do voice input anywhere, anytime, boosting your output efficiency.
06.16.2026 · 12 min
Gained 1000+ followers in a little over a day — here's what I did
In June I ran my X account seriously and gained 1000+ followers in a little over a day. This post records what I did and the feedback I got.
X operationsContent growth
A scorching take:Starting a business with Codex todayis like GPU mining in 2010!
I feel like most people still understand less than 1% of Cloudflare Workers.Workers are cyber laborers.They do the dirty grunt work behind your website for you.Submitting forms, uploading files, calling AI, login validation, handling payment notifications, and the API of a small product — Workers does all of it.The most outrageous thing about Workers: free users get 100,000 requests per day without spending a cent.What does 100,000 even mean?Say your site has 1,000 users a day. Even if it runs nonstop 24 hours and someone visits every single second, that's only 86,400 visits a day — still not enough to blow through the quota.If you can actually burn through all that free quota...what are you waiting for? You're about to get rich.Right now you can write the logic and just hand it to a Worker to run.
Pair it with the Hono API framework and that's the best practice.This is what makes Cloudflare Workers so delightful.Pay? Yeah, right.Cloudflare has a kind heart.It can't bear to see you spend money.
The fastest way to grow followers on X is to keep making people feel:not following you is their loss.Make yourself someone who's continuously useful to others.
On X, you must have opinions.People without opinions are just manufacturing information noise.With a stance, people remember you.With insight, people want to follow you.X has no shortage of re-posters — it's short on people who can explain things all the way through.
Sharing a small tip for posting on X:every post must stand on its own.Because you're not the center of the world — no one is going to click into your profile to piece together the full story.The first time a stranger scrolls onto you, they should instantly get a take, or a piece of know-how.Those are the only posts that carry value.
I am a dogged disciple of Wang Xiaobo.He's witty, but never slick.
He's funny, but always logical.
He loathed stupidity — and he loathed solemn, earnest stupidity just as much.If Wang Xiaobo lived in the AI era,he might say:finally, something that lets ordinary people do a little less dumb labor and keeps a little more time for genuinely interesting thoughts.That's also why I like AI.Not because it's divine.But because it can spare people a bit of the torment of tedious work.
In the AI era, the most important ability is meta-skills.In plain terms: the foundational abilities that still work when you switch tools, switch models, or switch industries.
- Can you state a problem clearly?
- Can you make sense of a pile of messy material?
- Can you break a vague idea into steps?
- Can you tell when AI is right and when it's bullshitting?
- Can you go back and forth with AI, turning a rough answer into something workable?
06.15.2026 · 6 min
2026 indie dev best practices: the Cloudflare "broke-ass all-in-one" stack
The zero-cost tech stack for indie developers: Codex for writing code, GitHub for version control, Stripe for payments, TanStack Start for frontend, Hono + Workers for backend, D1 for database, R2 for storage, KV for caching — all running on Cloudflare.
CloudflareIndie developmentTech stack
06.15.2026 · 8 min
How free users can squeeze Cloudflare dry — how far does the free tier actually get you?
Cloudflare's free tier can carry an entire personal internet infrastructure: DNS, CDN, Pages, Workers, KV, D1, R2, email, Tunnel, AI Gateway and more.
CloudflareFree tierDeployment
I feel like we've still developed less than 1% of Cloudflare.It has shipped way too many features.
- Object storage with R2
- Backend API with Workers
- AI gateway with AI Gateway
- Heavy compute with Container
- Caching with KV
- Database with D1
- Using PostgreSQL? Hook it up with HyperDrive