Ari’s Space*   est. 2026
Skip to content
ari*s space[ a. ]
← Librarythoughts and scribbles / 2026-05-02

Stop Prompting. Start Defining Outcomes.

Most people use AI like a slot machine. The makers I respect run a workshop. Three moves to flip your output, with the workflow that built Pushing Frames as a worked example.

9 parts / Hover to explore. Click to read.

Most people use AI like a slot machine. The makers I respect run a workshop. Three moves to flip your output, with the workflow that built Pushing Frames as a worked example.

Outcome + references
          |
     acceptance conditions
          |
        generate
          |
      compare + refine
          |
Save useful corrections
for the next session.
Give the model direction it can inspect, then judge the result against it. Saving a correction improves the available working context; it does not necessarily change the model itself.

I spent a year writing prompts at midnight, getting generic work back and blaming the model.

My briefs were often three lines long. I wanted the output to carry my taste, but I hadn't given the model enough to recognise it. Then I'd try a longer prompt or a different model and hope that would solve the missing direction.

What changed the work was getting clearer about what I wanted to make. The references, the audience, the decisions and the definition of a good result had to come from somewhere. I couldn't leave all of that in my head.

Write something you could hand to another person

Think about briefing a director of photography. You'd explain the audience, runtime, tone and references. You'd identify the frames that matter and the things the film must avoid.

Give an AI-assisted job that level of direction too. The useful question is whether someone could read the brief and understand what you would accept.

I learned this the irritating way on a carousel. I asked for “less wash” without specifying the overlay colour. Black wash and white wash need different changes. We spent a full render cycle on an interpretation I hadn't intended.

The instruction left the important part undecided. A more specific brief would have cost less than that evening did.

So write the result, the references and the acceptance conditions first. If the output misses, you can look at whether the brief was unclear, the implementation was wrong, or your understanding changed. Those need different fixes.

Give the model the context that is yours

Models can change. Your references, past work, voice rules and experience can stay with the project.

That is why I spend time on the workspace. I want it to contain examples of what I mean, rather than relying on a phrase like “make it feel premium”. What does that mean here? Which reference shows it? Which part of the reference matters?

Your context might include a decade in an industry, the jobs you've already made, or a distinction you notice immediately because you've spent years doing the work. The model doesn't automatically know which of those things should guide this project.

Brief stackMODELinterchangeablePROMPTcheapWORKSPACE & TOOLINGVOICE RULES, STYLE GUIDEREFERENCE LIBRARY, PAST WORK12 YEARS / INDUSTRY KNOWLEDGEYOURMOAT
Models are interchangeable. Context is yours.

In my own work, supplying the missing context has been more useful than endlessly polishing a prompt. That doesn't make the prompt irrelevant. It gives the prompt something specific to act on.

Build a small reference library you can explain. Add the decisions that keep coming up. Keep the examples that show your taste, including the ones you rejected and why.

Spend the saved time on the finish

I think of the change in craft as a 90/10 flip. It's a sketch of the shift, not a measured productivity promise.

The first usable version can arrive much sooner. The work then moves into deciding what's wrong with it: which frame is weak, which sentence sounds borrowed, which transition nearly works but doesn't quite land.

Old craft

GRIND TO 90%POLISH

New craft

AI HANDS YOU 90%TRIM PASS / TASTE / POLISH
this is where your craft lives now
The grind compresses. The polish expands. Your taste is the bottleneck now.

That last pass needs attention. You have to see the problem before you can ask for a useful correction.

Look at more work. Compare your output with the references. Cut the part you're keeping only because it took a long time to make. Use the time you get back to practise judgement, rather than treating a quick first version as a finished one.

Shape the workspace around your craft

My work is in video and stills, so I want the workspace to understand a campaign brief, find the references, and distinguish how I write for LinkedIn from how I write for Skool.

I built that context piece by piece. When something went wrong, I looked for the information that would have prevented it and added that to the working material.

For a photographer, that may mean references for light. For a writer, examples of rhythm and tone. For a director, a visual language the next session can actually inspect.

The useful workspace extends what you already know. You still need to judge what comes back.

Pushing Frames is a worked example

Pushing Frames is the image-generation plugin we built at Pushing Squares. The aim was repeatable frames that belonged to a campaign brief, rather than a collection of impressive but unrelated demo images.

What sits behind the plugin matters: the PS voice spec, CxN's reference library across stills and film, and twelve years of on-set and campaign experience.

Brief stackAI MODELreplaceablePUSHING FRAMES PLUGINPS VOICE SPECCxN REFERENCES (decades, stills + film)12 YEARS ON SET, CAMPAIGNS, VISUAL LANGUAGEWHATMAKESTHISNOTGENERIC
The model is the lightest layer. The context underneath is the heavy bit.

That experience includes things which are awkward to fit into a single instruction. How overcast light reads on skin. When negative space helps a campaign frame and when the frame simply feels empty. Whether a reference offers real direction or just looks good on a mood board.

Those are judgement calls. The brief is how you make them available to the tool. The model doesn't inherit your eye just because you have access to it.

The plugin itself was built the same way. The first version was rough. Its brief included the reference frames I wanted, the voice it should use to describe itself, and a definition of what “on-brief” meant. We used each version, corrected it and tightened the instructions for the next one.

Flow loop diagramDEFINE OUTCOMEwhat does a good frame look like?TIGHT BRIEFindustry context + style guide + voice rulesITERATE WITH TOOLbuild the next version of the toolSHIP FRAMESoutputs feed back into the brief
The tool that helps you make tools was built using the method it teaches.

The output fed back into the tool's design. You can inspect the repository and its briefs to see that process. The useful lesson is to let actual results change the next version of the working agreement.

Make the thing your work keeps asking for

An existing tool carries somebody else's defaults. Often those are perfectly useful. Sometimes you keep working around the same assumption because it doesn't fit your job.

That repeated workaround is worth looking at. You may be able to make a small tool around the way you actually work.

Pushing Frames came from a creative director's needs. Aris Space came from wanting a place to share knowledge that didn't feel like another Substack. The Pushing Squares plugins come from people who shoot, edit and deliver work.

You don't have to begin as a software engineer or start with a polished product. You do need a clear outcome, a brief and a willingness to try version one, see what's missing and improve it.

Start with the part of the job you understand well enough to judge.

Keep what the corrections teach you

I use the analogy of working with a junior: explain what worked, show what missed, and improve the instructions instead of expecting mind-reading on day one.

For an AI workspace, that improvement needs to be saved. A correction in a chat isn't the same as a durable rule or reference available to the next session. You're improving the context around the model; you aren't necessarily changing the model itself.

Over time, the workspace should need fewer repeated explanations. If it doesn't, check whether the examples and decisions are actually being kept and retrieved.

You still make the calls that need your judgement. The point is to spend less time explaining the same background before you can make them.

Try it on the next piece of work

Before you generate, write what the output is for, choose references that show what you mean, and describe the test it has to pass. After you generate, compare it with that brief and make the finish deliberate.

When a correction teaches you something useful, save it where the next session will find it.

That's how the work begins to look like yours. You have given it your direction, then stayed involved long enough to make the result meet it.

[ comments ]

guests welcome. members show first in the list

guests

  • 'Taste is now the bottleneck' and the 90-10 inversion. Two great insights to remember.

    guest · Constantine · Sun, 27 Sep 2026 17:08:21 GMT

1 / 11