Form is the medium — it decides what users and readers actually see. On WeChat, people look at the image and read the post body; on Video Accounts, people watch the video and largely skip the text.

Different platforms also mean different things behind the scenes: different moderation rules, different traffic distribution, different audiences.

It is only after I started making content for commercialization and product launches that these ideas became concrete. Before, I knew the big concepts — but did not really know them.

知行合一 (the unity of knowing and doing) — it feels something like this.

These days, building a 50-point demo with AI may only take half a day, or even an hour or two. But from 50 to 80, there is often a huge chasm.

Most of the time, shipping a 50-point demo does not prove much. Shortening the distance from 50 to 80 — delivering a product experience and service with the highest completeness you can — is what matters.

Of course, if you have already found your target users and a business model with a 50-point product, that is also fine. Knowing exactly what level of product you are aiming for matters a lot.

Image generated with GPT-image-2.

Vibe coding has a common detour: you ship the demo first, skip the thinking and research along the way, and just keep asking the AI to tweak it.

But sometimes it pays to stop and look at how reference products solve the same thing — how others handle the same kind of problem. It saves you a lot of dead ends.

Written after hitting a few of these while building ColorWander.

Two recent thoughts:

1. Ideas are easy to copy. Completeness and details are not. Marketing and distribution channels are not. The data and corpus you have accumulated cannot be caught up with overnight either. Spinning up a demo has never been easier — everyone can say "I just vibe-coded a XYZ." But shipping a product with a complete experience is hard. Getting 100, 1,000, or more people to actually use what you offer is hard.

2. A friend told me: "All the friends I know who used to build productivity tools — none of them want to build pure productivity anymore." I said: "I think it is a stage thing. The more optimistic you are about model and agent capabilities, the deeper you go, the more pessimistic you become about pure productivity, and the more you care about things that are playful, fun, and human." I genuinely believe that.

Building products that are fun and humane matters more and more.

I have been moving some of my dev tasks onto Codex lately, and I keep feeling it: a complete toolchain (calling image-gen APIs directly, reading markdown natively, clean integrations with all kinds of tools), a reliable user experience (highly visual), plus a GPT-5.5 brain that is back on its feet — it is genuinely good to use.

Each player leads for a month or two. OpenAI is back. No one gets to say they will lead forever, and never underestimate a strong player's determination to take a market.

As a tool user in this era, there is no good reason to lock yourself into one tool out of loyalty or habit. Whatever works best, whatever is more advanced — just use it.

One more thing: avoid free tiers and cheap substitutes — cheap things are always the most expensive.