Meta-building: how teas.co.uk was built with AI, one checked step at a time

Why a tea shop cares: one person built this tea shop with AI tools, and checking every step is what kept the facts about each tea right.

Building something with AI is fast. The risk is that it is also fast at going wrong quietly. Meta-building is the deliberate step back that stops that happening: after each piece of work, asking whether it did what was meant, not just whether it ran.

Metacognition, in plain words

Psychologists call noticing and checking your own thinking metacognition. It is an established word, and Lee does not claim it. Meta-building is the same habit applied to a build: noticing what you assumed, checking it against what is really there, and changing course when the two do not match.

The loop

1. Build one piece.
2. Inspect what actually came out, on the live site, not what the tool says it did.
3. Explain any gap: why is this value missing, why does this page disagree with that one?
4. Correct the source of the problem, not just the symptom.
5. Record the new state, the evidence and what is still unknown in a handover note.

Lee describes the loop, and what goes in a handover note, in How I build with AI: the loop, not the vibe.

A shop example

The instruction is "add decaf to the product filters". A quick build adds the filter and moves on. Meta-building keeps going. Which products does the filter now show? Are any decaffeinated teas missing, and if so, why: is the caffeine status blank in their records, or recorded in a different way? Does the filter keep decaffeinated and caffeine free apart, as the shop's decaf versus caffeine-free guide says it should? Does the live page show the same answer as the product records? Then the handover note says what was changed, what was checked, which records still need a caffeine status and what to do next.

The example is illustrative. The point is the order of work: data first, then the rendered result, then the written record.

Why it matters for a shop

A shop that asks people and AI assistants to trust its facts has to know when its own facts have drifted. Meta-building is the working habit behind the shop's single source of truth and behind the rule of built versus planned: you cannot say honestly what is built unless you have checked it.

What it is not

• Not a claim that AI does the thinking. The person whose name is on the shop stays responsible for the result.
• Not a guarantee. Checks miss things, which is why a failed check is written down, not hidden.
• Not a new science. It is a name for a careful working habit that many good builders already have.

In Lee's words: how I build with AI, the loop and not the vibe

This part is Lee Samuel Tucker's own account, first published as a news story on 15 November 2026 and moved here so the subject has one home.

Collins made vibe coding its word of the year. It is a good doorway. It is not how teas.co.uk was built, and here is the method that was.

In November 2025 Collins named vibe coding its word of the year: describing what you want in plain words and letting AI write the code. I understand why people would call what I do vibe coding. It is a useful doorway phrase. It does not describe how this shop was built.

Eighteen months of the loop

teas.co.uk took about eighteen months of building, stepping back, finding where things had drifted, meaning two copies of the same fact had stopped agreeing, writing down exactly where we stood, correcting it and going again. Vibe coding gets you moving. The loop is what stops the thing falling apart the next morning:

1. Build one piece.
2. Inspect what actually came out, on the live site, not what the machine says it did.
3. Write down the new state in a handover note, a short written account of what changed and where things stand, so the next session starts from the truth.
4. Go again.

Every change to this site still runs that way, including the Tea Wiki of 1,937 entries and the 402 products from 40 brands in the shop. The handover notes are the most valuable files I own.

What goes in a handover note

Mine follow the same six headings every time, so the next session, whether that is me tomorrow or an AI agent starting cold, can pick up exactly where the last one stopped. Here is a made-up example, the kind of small correction the shop deals with every week:

• Purpose: a box of tea was showing the wrong weight on one page.
• What I found: the product page said 500g; one of the feeds read by machines said 250g.
• What changed: the feed now takes its weight from the one product record, like the page does.
• What was checked: the page and the feed both read 500g. The price per cup in the feed did not update on the first check, so that check failed.
• Still not known: whether any other feed holds the old figure.
• Next: refresh the feed, check the price per cup again, then check the other feeds.

The failed check goes in the note, not just the fix. What I saw and what I plan to do are kept apart. And nothing private goes in: no customer details, no passwords, just the state of the work.

Several AI models, on purpose

I do not use one AI. I use several AI models, the systems that do the reading and writing behind an AI assistant, often on the same problem at once. What I am after is not answers. It is argument. One model sees the structure. One finds the gap. One catches a claim too strong to stand behind. One flatters me and is wrong, and I have to be sharp enough to notice.

The rule I took from that is on my founder page: check the machine's work. The responsibility stays with the person whose name is on the shop, which is why the site keeps built and planned firmly apart.

Paper first

I plan on paper, on purpose, even for the parts machines will read. Three lists before I touch a computer:

• the questions customers actually ask;
• the proof a stranger would need before believing us;
• the things nobody has bothered to write down that only someone in the trade would know.

Those lists became the product facts, the claim boundaries, the complete payload an assistant can ask for, and the Tea Wiki.

What it takes

The real cost is time: months of evenings before anything is good. One early job, rewriting roughly 300 product pages with AI and checking every suggestion myself, took about four weeks. The machines make the work faster. They do not make it shorter than it needs to be.

What I would tell anyone starting

Start with the thing you know better than most people alive, not the technology. Keep a written record of where you are. Ask more than one machine, and believe none of them until you have checked. Then publish what you can prove. Your "only" might be the whole trade.

Where the term comes from

Lee uses the phrase in his 2026 writing about how teas.co.uk was rebuilt. It is a name for a practice, not a claim that nobody worked this way before.

Quick reference

Related

• How I build with AI: the loop, not the vibe
• Single source of truth
• Built versus planned
• Lee Samuel Tucker's builder vocabulary

Download as PDF