Sketch-first, AI-second: how I used “visual assistance sketching” with Nana Banana to regain control in concept development
The last few days I’ve been doing a small but telling concept development project for my wife’s creative studio in Copenhagen. We’ve talked for a while about adding a greenhouse in the courtyard beside the studio — not as a “pretty add-on”, but as a spatial extension that could host guests, plants, and small gatherings.
The problem wasn’t imagination. It was control.
Like many designers, we started with the simplest thing: upload a reference photo and describe the greenhouse in words. And again and again, the AI produced something “nice”… but wrong. Wrong orientation. Wrong style. Wrong size. Sometimes all three. The outputs were visually convincing, but conceptually untrustworthy — and that’s a dangerous combo when you’re trying to make real decisions.
This post is about the shift that changed everything: moving from prompt-first to sketch-assisted prompting — and treating AI as a rendering collaborator that responds to your intent, rather than a generator that makes decisions for you.
Why pure prompting breaks down in precise concept work
Text prompting is great for mood and broad direction. It’s much weaker at:
- Geometric intent (exact footprint, alignment, setbacks, “do not protrude past this line”)
- Contextual constraints (existing facades, window rhythms, neighboring buildings, boundaries)
- Design continuity (matching the logic of the site rather than inventing a new one)
In our case: the greenhouse kept protruding into the parking lot, or drifting away from the gable, or becoming a different typology than what we wanted. I could tweak prompts forever, but I was essentially negotiating with a black box.
So I changed the interface.
The following sketch were my own to show the AI how we precisely wanted to locate the green house in the same plane as the Taverna facade, and this was a break through since we had tried so many times with failed attempts from the AI as if it was impossible to get the green house placed correctly
The workflow that worked: reference photo + your linework + AI rendering
1) Establish intent with a fast line sketch (not “pretty”, just precise)
I took the reference photo and did a simple pen sketch on paper — basically “layout geometry” in line form. The point wasn’t artistry. The point was constraint: where the new volume begins, ends, and how it relates to the existing building.
That gave me the first “sketch-mode” renderings. Still not correct — but noticeably closer.
2) Notice the real issue: the AI wasn’t “listening” to layout
Even with the line sketch, the AI still made an unacceptable move: it pushed the greenhouse into the parking lot. This wasn’t a minor styling drift — it broke the concept.
This is the moment many people give up, or they accept the wrong output because it “looks good”.
Instead, I escalated the clarity of the visual instructions.
The many failed attempts to get the green house placed correctly – and this is how I have seen most of the GPT/AI attempts from friends and colleagues to get the conceptual sketching to work using prompt only approaches – randomely placed concetps with little to no progress from round to round by the GPT
3) Add colour to signal hierarchy and material intent
Only when I added watercolor (simple blocks of colour) did the model start aligning with my intention: black profiles, glazed volume, warm interior light, planting density. The coloured sketch acted like a higher-bandwidth briefing.
At this stage, moving to the Nana Banana (Google) model was a turning point — it responded better to the visual logic and produced outputs that were structurally closer to what we meant.
The first attempts to place a green house in front of the house. I think the Neon Light signage was making the AI place the green house further left, even though we tried a lot to get it to fall even with the house facade. Each of the attempts look good in itself but fails to meet the idea .
The Failed Prompt:
Make a rendering where there is a black wooden green house with a double door filling out the space between the two buildings. There is still
The taverna Centrale laminate the greenhouse and we can see it through the windows. Also there is a cafe on the ground floor and the greenhouse house doors are open so we can see the cafe life inside the greenhouse house doors. And there is a lot of green plants backlit inside the greenhouse and we can see th green plants and bright cozy light through the windows.
Outside there are planters with mature plants and poles with light strip as shown on the sketch
The greenhouse facade sits in the same plane as the taverna building.
4) Use “draw-over” on the best
AI result to correct what it still gets wrong
Once I had an output that was close-but-not-quite, I stopped trying to fix it with text.
I opened the photo and used iPhone sketch mode with ruler + white lines to draw the true boundary and roof direction. Then I asked Nana Banana to fill in the space according to those marks.
That was the “voilà” moment: the first acceptable greenhouse concept, correctly placed and behaving like part of the building rather than a decorative object.
(The rendering of the house seemed right with the good style and vibe, however the building was too far from the brick building which is our studio and where we want it to be adjacent to. Therefore I sketched the white lines on top to help the prompting to extend the green house, later I introduced a “1 m gap with a security gate” to the prompt as well to create some distance to the white building
The successful prompt:
It is almost right but you have to connect the green house with the taverna Centrale building like the white lines I have added so you can see what I’m trying to do.
That is great. Now make the full green house with same finish inside so have the lights and tables fill the whole house and put the same weight on the black profiles.
Make the planters stand outside and in front of the glass house and have the light chain continue hanging out in foreground in front on to some poles that are fitted to the planters. of the green house and attach to the wall of the white building in front like on the reference photo
5) Iterate conceptually: now AI becomes fast, obedient production
With the core geometry nailed, iteration became easy:
- The studio facade in reality has 8+ windows, not 3 → I asked for the longer version
- I asked for a brighter daytime version with more transparency and visible guests
- Finally, I asked for a simple 2-second live-view video to communicate the vibe
Total time spent: around two hours — including sketching, photographing, draw-overs, and prompt iteration. But critically: those were two hours of steering, not browsing random alternatives.
(The two different visual pathways I followed, the left building is more modern and the frames don’t really fit our 50’s creative vibe, whereas the one to the right is more old and simple with lighter frames and seem to fit the building well)
(the final picture that has most of the style I want to share for the concept description for the cafe and greenhouse. the style is lighter and I like we added people and green plants to the house. I have not yet added the Taverna Centrale red signature branding on the brick building, but for the conceptual push – I think this is suficient when discussing the idea further
The bigger point: we must keep our sketching skills alive
This is where I’ll be blunt as a working designer and concept developer:
If we outsource intent to AI, we will slowly lose the muscle that makes design… design.
Not because AI is “bad”, but because it is opinionated by default. It fills gaps. It improvises. It optimises for plausibility. And when you’re doing conceptual work for real spaces, real products, or real decisions, plausibility is not enough.
Sketching — even rough, even “ugly” — is how we:
- externalise intent
- define constraints
- choose what matters
- stay in authorship
The goal isn’t to compete with AI’s rendering. The goal is to direct it.
(the video was generated by Nano Banana as well with the intent to showcase how an afternoon with the new Green house could look like. The style is still the modern looking direction that I later dropped, but the animation it self is really key to undersatnding such concept and is practically free once the visual has been made)
A practical principle: left footprint, right footprint
The best way I can describe the collaboration is like walking:
- Left footprint (human): define intent with sketch, boundary, hierarchy, constraint
- Right footprint (AI): render, vary, enrich, light, populate, materialise
- Repeat.
When you work this way, AI becomes a creative assistant with great rendering chops — not a creative director that decides your concept for you.
If you want to try this tomorrow: a simple recipe
- Start with a real photo (the context is your truth anchor)
- Do a fast line sketch over it (paper or digital) to lock geometry
- Generate 3–6 variations (don’t chase perfection yet)
- Pick the closest one and do a draw-over correction with ruler lines
- Regenerate with the draw-over as the new constraint layer
- Only then: style passes (day/night, planting density, people, signage, materials)
Leave a Reply