Working prototype · Hadi Al-Hazim · August 2026

Design Direction
Explorer

Divergence fuel for the first hour of concept work. Every image here is a provocation, not a deliverable.

A designer's complaint about AI is not that it is slow. It is that what comes back feels generic, so it makes them a faster designer rather than a more imaginative one. Generating three images from one prompt does not fix that: it returns three versions of the same idea. This runs three fixed stances against the same brief instead, so the outputs disagree with each other. Each one also says what it steals from the inspiration and, more usefully, what it refuses.

Stance 01Amplify the inspiration
Stance 02Invert its logic
Stance 03Cross it with another climate

These three stances were written by AI, not by a designer. They are a starting point to argue with. In use the designer writes the stances, because choosing what to argue against is the judgement, and it is theirs. The toggles below are a different matter: those are lifted from the firm's own published vocabulary.

Future state

Today it argues from nothing.
It should argue from your own back catalogue.

The prototype above works off one synthetic image and three stances an AI wrote. That is enough to test whether arguing with a designer is useful at all. It is not enough to be worth using twice. Here is what changes, and what stays honestly missing.

Now

The prototype

InputA brief, three toggles, and one synthetic inspiration image
↓
Stances — written by AIAmplify · Invert · Cross. Fixed, generic, the same three for every project
↓
OutputThree directions that disagree, each stating what it steals and what it rejects

Honestly missing

  • No eval. Nothing here can tell you whether a direction is any good.
  • Knows nothing about the firm, its projects, or what it already tends to do.
  • The stances are not the designer's, so the argument is not theirs either.
Next

With the firm's own record

InputThe live brief, and the designer's own reference images
↓
Stances — written by the designerThey choose what to argue against. That choice is the judgement, and it belongs to them
↓
Retrieved, not trainedAn index of completed projects, material palettes and post-occupancy notes. Adding last month's project is an upload, not a retrain
↓
OutputDirections that argue against the house's own habits: "six of your last eight tech HQs put the social heart at the lift lobby — here is one that does not"
Feedback edge The designer marks which directions actually changed the drawing. That signal tunes retrieval and stance weighting, and becomes the eval the prototype does not have.

Why retrieval and not a tuned model

  • A model tuned on the firm's work is frozen at the day it was tuned, and moving it costs a retrain.
  • Worse for this brief: it would learn the house style and reproduce it. That is a generic-output machine with better manners.
  • Retrieval keeps the catalogue current as a content job, and keeps the firm's judgement as the asset rather than the model's weights.
Enlarged concept visual

  ·  click anywhere to close