Anatomy of one build.

Input 1 · the client's truth

The brief.

This is the complete client file the engine received — the stand-in for an answered intake form. Rafid is fictional, invented to test the pipeline; his placeholders are marked as placeholders so the engine can't mistake them for facts.

An honest note on this artifact: the brief's shape is provisional. AURO's real intake will be Galen's layered questionnaires, and this file will converge toward whatever those produce. The questions themselves are part of the design surface — every field here either earned its place in the output or didn't, and that feedback should flow back into the intake form.

Input 2 · the taste

The skill.

The instruction file compiled from Galen's words — thirteen hard rules, a direction register, and a self-check floor, sitting on a general craft layer. It travels into the build as a file beside the brief; the engine reads it before writing a line of code. Read it rendered, or raw, exactly as the engine sees it.

Input 3 · the instruction

The prompt — all of it.

This is the entire instruction the engine was given, recorded verbatim in the build's manifest. It is deliberately small: all the taste lives in the skill, all the client truth lives in the brief. The prompt only connects them.

Read SKILL.md in this directory and follow it exactly to design and build the website for the client brief in brief.md. Write the site into ./site/ and the sidecar documents beside it, per the skill's output contract. Use file tools only: no shell commands, no web browsing. Work until the output contract is complete.

The machine

Build conditions.

Recorded per build in a manifest, so any two builds can be compared honestly — same skill or different, same brief or different, and nothing else quietly changing in between.

• modelClaude Opus (claude-opus-5), run headless — no human touched the output
• toolsfile reading and writing only: no shell, no web browsing, no image generation
• workspacea fresh, empty directory containing exactly three things: the brief, the skill, the craft base
• fingerprintsskill, brief and craft base each carry a checksum in the manifest — “which version ran” is never a guess
• variancethe same inputs also ran a second time (take B on the wall) — the gap between the two takes shows how much is skill and how much is chance
• not yet recordedcost — time and tokens per build. Honest gap; see the open items below.

What came out

The paper trail.

Beside the site itself, the engine must ship its own working documents. For this build, all three, unedited:

• 01DESIGN-SYSTEM — the named intention (“Eerst uitleggen, dan pas sleutelen” — explain first, then wrench) and every token derived from it, palette to motion
• 02ASSUMPTIONS — every gap in the brief and what the engine did about it: hours, prices and the company number left as visible placeholders, nothing guessed into the code
• 03SELF-CHECK — the engine grading itself against the floor. Read it with one caveat: this round's independent checkers caught self-checks over-claiming, which is why external grading exists.

And the site itself, shown on the wall exactly as delivered — flaws included.

Humility section

Deliberately unsolved.

Not a roadmap, and not complete — just the open ends both of us can see from here.

• Cost. These builds ran on a flat subscription, so the price of one site is not yet a number. Moving generation onto the metered API — and recording time and tokens in every manifest — is what turns “an engine” into “a unit economics answer.”
• The intake form. The brief above is a stand-in. Converging it with the real layered questionnaires — and letting each build's assumptions log tell us which questions were missing — is open work.
• Photos and imagery. The single biggest untested area of taste: what happens when a client does have photos, or wants stock, is not yet in the rules.
• The working rhythm. How many directions per client, who reacts when, how corrections loop — proposed in Q4 on the showroom, waiting for a yes, a no, or an adjustment.
• A real client. Everything on the wall is fictional by design. The first real brief will teach more than any test run.
• Sharper checks. Two mobile defects shipped past this round's checks (they're visible on the wall). The fix is mechanical — screenshots at phone width inside every build — and staged for the next round.