OpenClaw rebuilding my family's farm website - Here's what surprised me

I did most of this project one-handed.

My son Mino is five months old. He needs soothing to sleep — rocking, patting, slow pacing through a dim room. My other hand had my phone open to Telegram. That’s how riofarm.vn got rebuilt over four days: my sister-in-law’s macadamia farm website, migrated from a slow Gatsby template to Astro 5, with a full content strategy, nine products, twelve blog posts, and a working cart.

All managed via chat. No laptop. No IDE. No “let me sit down and focus.”

I’m a product manager, not a developer. I’ve used AI coding tools. This felt different — and I want to be specific about why.

---

The Ask

I wanted a complete rewrite: Gatsby to Astro 5, zero feature losses, full test coverage before shipping, zero downtime on the live domain. I gave the brief over Telegram:

“Write the tests before you code. Tests must cover 100% of code paths. Leverage unit tests to maximize test coverage but don’t forget e2e UI tests (playwright). Old gatsby code must be kept intact locally as a backup. I don’t want any downtime.”

The agent responded that it would spawn a Claude Code sub-agent with a comprehensive brief, then check back on progress. Twenty-one minutes later it pinged back: 62 Playwright E2E tests passing, 29 unit tests passing, Astro site built and deployed to Netlify preview.

The sub-agent model is the right mental frame for this kind of work: you write a brief, a coding agent takes it, runs independently for 20–60 minutes, and delivers a reviewable output. Not instant — closer to handing off to a contractor. But you can go put your baby to sleep in the meantime.

---

What Surprised Me: The Agent as Collaborator

Here’s where expectations broke down — in a good way.

After the core migration, I asked a simple question: “Help me enrich the content of the astro site. What more info should be displayed there? Be mindful that it’s a Vietnamese site tailor fit to Vietnamese consumers.”

I expected content suggestions. What I got was an unsolicited gap analysis, prioritized by business impact:

Note that the agent couldn’t even load the Facebook page I’d linked (login wall). It worked from the site alone and the words “Vietnamese consumers.” The response ranked what was missing by what would actually drive sales:

1. Missing product pages (dried mango, dried jackfruit weren’t on the site)
2. Zalo CTA — “Vietnamese consumers use Zalo more than any other channel. The site only shows a phone number buried in the footer. A sticky Zalo button would meaningfully increase conversions.”
3. Customer testimonials — “Social proof is critical for Vietnamese e-commerce. ‘Khách hàng nói gì’ section with authentic quotes from customers.”

The original Gatsby site had no Zalo integration anywhere. I hadn’t mentioned Zalo once. It connected those dots from four words: “tailor fit to Vietnamese consumers.”

That was the moment I stopped thinking of this as a coding tool. The reasoning chain — Vietnamese site → Zalo is the primary channel → phone number buried in footer → sticky button would increase conversions — is product thinking. It’s the kind of gap analysis a good PM does before touching a spec. It happened before I asked for it, while I was in another room putting my son to sleep.

On testimonials: the initial proposal was for placeholder quotes. I had screenshots of actual customer chat messages and sent them instead. The agent read them, extracted the quotes verbatim, styled a testimonials grid with the farm’s warm yellow palette, and published it. The quotes stayed as they were — casual Vietnamese, emojis intact (“Mít ngon lắm em. Bọn nhà chị ăn hết 2 cân rồi kk 😂”). No sanitizing, no corporate polish.

---

What Surprised Me: Autonomous Execution

The other surprise was what happened when I handed off research tasks.

The farm’s homepage had an embedded YouTube documentary — a local news segment about Rio Macca. I asked for blog posts using facts from it. The constraint: don’t fabricate anything — if the data isn’t in the source, flag it.

Instead of watching the video, the agent ran a script:

from youtube_transcript_api import YouTubeTranscriptApi
transcript = api.fetch('BUKddHPS3pk', languages=['vi'])

YouTube auto-generates Vietnamese captions. Full transcript extracted in seconds — direct quotes, production details, the founder’s story, all pulled programmatically and cited. Three blog posts written from source material, no browser, no manual reading.

Then, cross-referencing a Báo Lâm Đồng newspaper article I’d shared, it found facts the documentary didn’t have: the farm’s output had grown to 20 tons/year (the documentary said 10 — the farm had grown since filming), and an OCOP 3-star certification — a Vietnamese government quality mark — that wasn’t anywhere on the site. Within a few minutes, the 10 tấn references across three blog posts were corrected to 20 tấn, the OCOP certification was added to the founder story, production process post, and homepage, and the generic “VSATP certified” trust badge became ”🏅 OCOP 3 Sao.” One cross-reference. Zero manual diffs.

When I asked about harvest season months — not in either source — it asked me directly rather than filling in a plausible answer. I didn’t have the data either, so I said to skip it and work from what existed. It did.

That disposition — knowing when to ask instead of guess — matters more than it sounds. For a small family business where credibility is everything, a confident wrong answer is worse than no answer.

---

---

What the Tests Don’t Catch

The test suite passed before domain cutover — 62 E2E, 29 unit, all green. They covered cart flows, routing, form validation, mobile nav.

What they didn’t catch: the Zalo icon rendering as a plain letter “Z.” The Zalo button covering the cart checkout when the drawer was open. The checkout button scrolling off-screen on mobile. A CTA pointing to a 404. All four found by opening the site on a phone.

Tests cover the logic you specify in advance. They don’t replace someone actually using the product on the device your users have. The fix for each was minutes once reported — but the finding required a human.

---

What This Means for PMs

The honest capability is more specific than “AI builds your product while you sleep.”

It’s this: if you can write a clear brief and define what done looks like, you can delegate substantial technical work to a background process and return to something reviewable — from your phone, between other things. The shift is from writing code to writing briefs. The constraint that doesn’t change is judgment: someone still has to look at the result, use the product, and catch what the tests miss.

What’s new is the collaborator dimension — an agent that reads context, identifies what’s missing, and proposes it before being asked. That’s not autocomplete. It’s closer to a junior hire who actually thinks about the problem.

The riofarm.vn project ran Feb 16–19. Live site, real users, real products. Not a demo.

If you’re a PM who hasn’t yet built something real with an AI agent — not asked it questions, but handed it a brief and reviewed the output — the gap between you and someone who has is wider than it looks from the outside.

The shift is hard to describe until you feel it. Pick something small and real, and find out.

---

OpenClaw is an open-source self-hosted AI gateway. riofarm.vn is live.

---