TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · The Brief · AGE 16—18 · POSITIONAL · METHOD
Iterate like an experiment
A brief revised in five places at once gives you a new result and no knowledge.
When to use
Every time an output is wrong and you are about to revise the brief. Especially when you are tempted to rewrite the whole thing.
The method
Two operations, and they answer different questions. Diagnosis: keep the brief that produced the result, copy it, change one thing — one constraint, one reference, one word of intent — and run it. This tells you what the brief is sensitive to. It does not prove the change caused the difference, because a generative tool is stochastic: the same brief, run twice, can differ. Comparison: hold the model, its version and every setting constant; fix the seed where the tool allows; run each version several times, not once; and compare the tendencies across the runs, not one image against another. Write down what you held, what you changed, and what moved. A brief revised in five places at once, run once, gives you a new result and no knowledge. One change, run once, gives you a suspicion. One change, run five times against five unchanged runs, gives you something you can say in crit.
Watch for this
Trusting one before-and-after pair. On image tools, the difference you see can come from the random draw rather than from your edit. And changing one thing at a time cannot show you when two things only work together — a material word that matters only once the light is set. When you suspect a pair, test it together as well as apart, and say so in the note.
The Lab's note
Isolate to diagnose, repeat to compare: this is the Lab's discipline for briefing, carried over from the science class and corrected by it. Changing one factor at a time is the beginner's experiment, and it misses what factors do together; repeated runs with settings held are what let you say anything about a tool that answers differently each time. The Judgment strand's on-ramp card teaches the first half to younger students; this card adds the second.
Try it
Take one brief that gave a poor result. Make three copies. In each, change a different single thing: the intent line, one constraint, one reference image. Run each copy three times, and the original three times. Which change moved the result most, across all its runs? That is where the brief was weakest — and where one run alone would have misled you.
Prove it
Show a chain of at least four changes, each run more than once with the settings held, with the observation for each — and name the change that mattered most and the one that turned out to be the random draw.
How it works
This is the controlled comparison, applied to a stochastic tool. Holding everything constant and varying one factor is what lets you attribute a difference; repeating the run is what lets you see the difference through the noise. Generative tools make both cheap — a run costs seconds — and that cheapness is the trap, because it invites changing everything every time. The six-layer brief names the layers you can vary; the versioned brief keeps the record, including the model and settings that produced each output. Together they turn prompting from guesswork into something you can learn from.
What this idea builds on
- A starting idea.
What this idea opens up
Sources
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.