I've Joined the Vibe Coding Wave

Video

It’s 2026 — are you still typing code line by line?

I’ve completely fallen into the “Vibe Coding” rabbit hole. From web chat boxes to Cline, and now as a full Cursor Ultra subscriber, this isn’t just a tool upgrade — it’s an efficiency breakthrough of a different magnitude.

A Late Entry

I consider myself part of the later wave to enter Vibe Coding. Until the second half of 2025, I was only modifying code via web-based chat — confirming logic in the chat box and then manually copying it in for debugging. It hardly counted as “agent development.”

By the end of 2025, I started experimenting with the Cline + Poe API combination. Initially, I only tackled highly deterministic tasks, such as encapsulating interfaces or writing process scripts and documentation based on verified, working interfaces. During this process, given the wide variety of models on Poe, I gradually gained a tangible sense of the capability gaps between them — the Claude Sonnet/Opus series truly left the rest in the dust, making almost no mistakes; the few errors that occurred were, upon review, due to my carelessness in selecting the wrong model.

Why I Switched Away from Cline

However, I soon realized that Poe was not suitable for programming agent scenarios like Cline, mainly due to three pain points:

1. No Context Caching: Every request is recalculated from scratch, causing token bills to skyrocket for long tasks.
2. Poor Long-Connection Stability: Cline relies on persistent connections, but Poe has limited support in this area.
3. Complex Remote Development Scenarios: I frequently work by SSHing into servers. To ensure the VS Code Server on the target machine could stably access the proxy network, I had to fiddle with port forwarding or proxy configurations on every machine — extremely inconvenient.

For these reasons, I began looking for alternatives and eventually switched entirely to Cursor. It was practically born to solve the aforementioned problems, so I also subscribed to the Ultra membership.

What Vibe Coding Changed

With the boost from AI programming agents, I noticeably felt I had much more time on my hands:

• Paying Off Technical Debt: Interface encapsulations previously shelved due to time constraints were finally completed one by one, with demos and documentation updated simultaneously.
• Batch Experiment Scripts: I could have the AI systematically design and generate large numbers of experiment scripts, rather than writing them by hand one at a time.
• Faster Algorithm Deployment: Once algorithms and models were finalized, deploying Python prototype inference scripts to production was essentially deterministic, repetitive work. Especially since we had fixed the C++ inference interface specifications, adding new algorithms to the deployment end could almost be entirely delegated to the AI — which performed quite reliably.

Unexpected Toy Projects

Besides work tasks, I was also able to carve out time to try some side projects:

• Eliminating Malicious Mining Programs: Two machines in the lab had been infected with malicious mining processes that would respawn no matter how many times I manually killed them. I had the AI write an automated “health check” script to comprehensively track and clear suspicious processes — with significant results.
• Daily Tech Briefing Service: I built a service that automatically scrapes the latest tech news daily and generates summaries, in bilingual form, displayed on a dedicated website. Now I habitually scan it every morning; it has become part of my morning routine.

A Bit of Calm Reflection

Overall, entering the first half of 2026, AI Coding has become quite robust:

• For deterministic tasks, it almost never makes mistakes.
• For vague requirements (e.g., I don’t know much about frontend and can only describe the general framework and feel), it still delivers decent results.

That said, the ceiling of an AI agent is always limited by the user’s own understanding. While enjoying the efficiency gains AI brings, one must also continue to deepen their technical knowledge — don’t let your own blind spots become the bottleneck that holds the AI back.