My daughter asked me to draw her a shark. I can’t draw. What I can do is open Claude Code and describe what I need, so that’s what I did — and it turned into a small, ongoing project of making coloring pages for a 3-year-old.

The Need: A Blank Page, on Demand

Store-bought coloring books run out, and they’re never quite the right animal on the right day. My daughter wanted a shark family — a papa shark, a baby shark — right then, not the next time we happened to be at a bookstore. What I actually needed was a way to turn “draw me a shark” into a printable page in a few minutes, using tools I already had open.

The Proof: A Page She Can Finish

The pages that work now share three rules I didn’t know I needed until I got them wrong:

  • White background, colored outlines only. My first attempts had a light blue background behind the shark. It looked fine on screen and burned through printer ink fast. Now the rule is: background stays white, and each shape gets an outline in roughly its real color — a black outline for a hedgehog, orange for a shark’s mouth — with the inside left blank for her crayon.
  • Color-coded lines, not just black. A butterfly with five wings in five different crayon colors is a lot easier for a 3-year-old to follow if each wing’s outline is already that color. She matches the crayon to the line instead of guessing.
  • Print double-sided — one sheet, the same picture on both sides. Not two different pictures, one per side — the exact same one, twice, so one sheet of paper is two chances to color it. I asked for two print buttons, single-sided and double-sided, and clicking double-sided is the part that actually surprised me: I never touched my printer’s own settings, and the same picture still came out on both sides of the page.

None of that is exotic. All of it took an actual page in her hands, colored wrong or wasting ink, before I wrote it down as a rule instead of a guess.

The Story: What’s Still Not Solved

Not every page comes out clean on the first try. AI image tools are good at drawing a shark; they’re not consistently good at drawing only the outline of a shark, in the exact colors I ask for, with nothing shaded in between. I’ve re-run the same request three or four times to get a hedgehog that looks like the hedgehog in her actual stuffed animal instead of an abstract shape, and I still haven’t found a web tool that reliably does a clean, color-matched, white-interior outline on the first pass. I’m looking at whether writing my own small script beats leaning on general-purpose AI image apps for this one specific job — outline art is a narrow enough task that it might not need a full image model at all.

The same project also produces the reference art for her nursery-rhyme player’s illustrations, where getting a picture slightly wrong has already taught me a separate lesson about asking AI for one change at a time.

The How: A Small Folder, Not an App

There’s no real app here, on purpose — no login, no dashboard. Just a folder on my home server, one image per coloring page, with the reference photo I used to generate each one saved alongside it so I can check later what I actually asked for, plus a small print page with two buttons: single-sided and double-sided. The whole project cost nothing beyond ink and paper — the same printer I bought for scanning sermon booklets does the printing here too.

Step by Step

  1. Start from a real reference, not a description, whenever the subject is specific. “Draw a hedgehog” gives you a generic hedgehog. My daughter’s actual stuffed hedgehog toy needed a photo of the toy itself attached to the request — description alone kept drifting away from the real shape and colors.
  2. Ask for the constraint explicitly, every time, in the same words. My working request now always includes the same four conditions: white background, colored outline only, each part outlined in roughly its real color, and nothing shaded or filled in between the lines. Leaving any one of these out is what produces a page that looks nice on screen and wastes ink or can’t actually be colored.
  3. Check the result against the constraint before printing, not after. I look at the image for exactly two things: is the background pure white, and is the inside of every shape actually empty. If either fails, I ask again rather than accepting a “close enough” version — a shaded interior can’t be undone by printing lighter.
  4. Re-run with a narrower ask if the first result adds or removes things you didn’t ask for. This is the same lesson as asking AI for one change at a time — the more you ask for in one request, the more of the picture the model treats as fair game to redraw.
  5. Save the reference photo next to the finished image, in the same folder. Every coloring page sits beside whatever I used to generate it. Months later, when she asks for “the shark one” again, or a slightly different version, I don’t have to reconstruct what I originally asked for.
  6. Duplex is a printer default, not something you switch on per print. The image file itself has no double-sided setting; it’s always just one flat, single-sided picture. What actually puts it on both sides of one sheet is the printer’s own saved default, set once in its settings program rather than picked in the print dialog each time. My printer’s default was already double-sided, so clicking print was enough — there was no duplex box left to check.

💬 Prompt that worked “My 3-year-old wants a shark drawing. I can’t draw. Make a coloring page a dad could print at home — papa shark and baby shark, colored outlines only, white background, nothing shaded in, so she can finish it with crayons.”

It’s not a big project. It’s a folder that grows by one picture every time she asks for something new, which is most weeks — and a six-step routine simple enough that anyone could run it without knowing how to draw either.