So I typed one sentence: "draw the entire AIOS project." Out came a hand-sketched map of my whole operator brain — six live connections, one central hub. Then I looked closer at the labels.
Last episode I wired up a skill that sketches like a human. So I pointed it at the hardest thing I had: my whole AI operating system. No diagram tool, no boxes-and-arrows by hand — just one line, and a rule that every picture also saves an editable twin.
"Calendar 2/9942." "Email Ee03130." The image model had quietly written raw colour codes into the text. This is the exact part most AI demos crop out — so I'm leaving it dead centre.
When the styling leaks a stray colour like #Ee03130 into the prompt, the image model can't tell "this is a colour" from "this is a word" — so it draws the characters onto the label. The editable .excalidraw twin had clean text the whole time, and a 5-cent re-run fixes the PNG. The glitch is the lesson: you stay in the loop, the tool does the drawing.
The same command that drew the system map also drew the pieces inside it — brand palette, onboarding cards, the job-hunt pipeline. Every PNG ships with an editable twin, so a wrong label is never a dead end. Tap any image to enlarge it and read the exact prompt that drew it.
Same loop, run again — typed once, rendered through kie.ai, saved as a PNG and an editable file. The labels behave when the colour codes stay out of the text.
Grab the editable AIOS system map and the two slash commands that drew it. Open the .excalidraw, swap in your own boxes, and you've got a map of your stack. Everything's free and open.
gh repo clone bitesize-ai/episodes ~/bitesize && open ~/bitesize/ep02-system-map
No GitHub? Comment MAP on the post and the bot DMs you the link.
If the system is a map, the connections are levels to unlock. Next: I batch-generated 18 gamified "unlock cards" for my whole operator setup — from my phone, on the couch, in one Telegram session.