Field Notes · No. 01 · Building in public Live telemetry

The Digestion

Two days of shipping, seen from the platform floor — the deploys, the database, the backend, and the one question underneath all of it: is the machine honest about what it knows?

A 48-hour window · 2026-07-11 → 07-13 · vibecrafting.ai · read from live Vercel · Supabase · Railway · git
47Registered users
21Made a design
68Designs total
6New in 48h
9Changes shipped
0Backend errors

ALL FIGURES READ FROM LIVE TELEMETRY AT TIME OF WRITING · 0 CONFIRMED PHYSICAL BUILDS YET — THESE ARE MAKERS, NOT BUILDERS

00

The turn

before → passes → now
X Before

The model believed itself.

Describe a table and it would hand back a confident, finished-looking design and stamp it valid — genuinely believing you could build it. It wasn't lying. It just had no way to be sure.

Y The passes

We made it check its own work.

There are no builders yet — only makers and users, people who've made a design, not cut wood from one. So the correction didn't come from the bench. It came from the model auditing itself, pass after pass, iterating from each failure.

Z Now

A truer pass — for boxy furniture.

The rectilinear stuff, we can stand behind. Curves are still TBD — and the 25% honest number says exactly that: the shapes we can guarantee pass; everything we can't yet build is flagged, not waved through.

01

The one-line verdict

health

Green across the stack. Two paper cuts, one scaling watch.

Stable

Frontend, backend, and database are all healthy and live. In 48 hours the product shipped a full buildability arc — designs the tool honestly can't build get caught and repaired instead of waved through — plus a buildability benchmark, a matured workshop, and tapered legs. All merged, all deployed, zero production backend errors. The only real defects are cosmetic-tier, and the one thing worth watching is a scaling question, not a live one.

02

Platform floor

frontend · backend · database

Every light is green

Healthy
LIVEvibecrafting.ai
200every observed request
0backend errors / 48h
1.00db cache hit ratio

The frontend serves clean, the backend boots clean and returns 200 on every observed request, and the database lints clean with a perfect cache-hit ratio. Nothing in the window points at a live problem.

The honest watch-items are all scaling ones: heavier per-design payloads and a couple of admin queries that are invisible at this size but worth bounding before any real growth push. None of them are firing today — they're notes-to-self for the version of this with a hundred times the traffic.
A full deep security sweep is on the pre-growth checklist, not skipped — flagged here in the open so it doesn't get quietly forgotten.
03

The number that looks like a bug

read this before you panic
Valid designs17 / 68
Flagged not-buildable51 / 68
Made a design21 / 47
25% valid is the honesty, not the failure. The validator flags about three-quarters of the stored library — almost all of it early, free-form geometry from before the engineering rules tightened — as not buildable, rather than stamping a green check it can't defend. "Passes ✓" must mean "you can actually build this," or the checkmark is a lie. That's the entire thesis of these two days. The fix isn't loosening the validator to make the number prettier; it's template coverage — correct-by-construction geometry that passes because it can't not. Each new template pulls another slice of that 75% back into buildable — and honestly. Today that solid ground is boxy furniture; curves are the frontier just ahead.
04

What shipped in 48h

all merged · all live
◆Buildability arc — the tool now catches designs it can't honestly build (missing parts, impossible legs, floating fasteners) and repairs them, across 20 designs from 13 different makersengine
◆Buildability benchmark — the ruler. A fixed test corpus that measures what fraction actually builds, and ranks which template to build next by real demandeval
◆The Drafting Line workshop — warm redesign, shop-buddy voice, an honest decline when it can't build something, plus a dashboard and a Finish Wall for what you've madeux
◆Tapered legs — the geometry engine can now splay and taper a leg to a foot; the on-screen 3D render of it is the next pieceengine
◆The reveal — a trusted blueprint is now the default hero, and the old "don't fully trust this" disclaimer is gone because it no longer needs to be thereux
◆Finish Wall, persisted — the first time you export a build now gets remembered, so your shelf of finished work sticksfull-stack
◇Deferred: one dependency bump was a major-version jump that broke the build in preview — held for a proper migration rather than forced throughheld
05

Architecture state

generation ⊥ validation

Front 1 — Stock & joinery

Done

The half that generates a design now reads every stock and joint fact from the exact same source of truth the half that checks it reads from. A guarantee test proves there are zero "we generated something we can't source" failures across every builder. The two halves literally can't disagree about what wood exists.

Front 2 — Assembly closure

In progress

You can't select your way out of arbitrary geometry — some shapes just don't close into a buildable object. Templates fix that by construction: they guarantee a valid assembly because they're built to. The benchmark measures which shapes pass and ranks which template earns its slot next by demand. The splayed-leg frame is first up — the first real step past boxy, into angles.

The through-line: generation and validation each used to keep their own partial, hand-maintained copy of "how wood works" — and quietly drift apart. Every fix this sprint collapses that gap: the two brains become one. That's why the work reads as a single story, not eight unrelated tickets.
06

What's next

the ranked list
  1. engineBuild the splayed-leg template. It's the #1 demand-ranked shape and it's proven achievable — the first step past boxy furniture into angled geometry, and the next big slice of that 75% coming back. This sprint.
  2. uxRender the tapered leg on screen. The engine already understands the geometry; the visible half is still dark. Close it.
  3. hygieneSmall stuff, done right. Resolve the deferred dependency bump properly, and bound the admin queries before they ever matter.
  4. growthThe first real look at the funnel — and a full security sweep — before any growth push. Honesty scales the same way the validator does: measure before you promise.

Describe something you want to build. Watch it get an honest answer.

Try a sample →