All posts

Two AIs Draw Me

I asked the AI coding agent I work with every day to describe how it pictures me, as a prompt for an image model. It refused to invent a face, so the figure is seen from behind. Everything else in the prompt is real work we did together.

I pasted the prompt, unchanged, into ChatGPT and Gemini.

Prompt
A cinematic wide-angle digital illustration of an image-algorithms engineer and
technical lead working late at night, seen from behind at a sleek desk, so the
face is not visible. The person wears a dark hoodie and over-ear headphones with
a violet accent, rim-lit by cyan and violet neon light. Three glass-like
monitors curve around the desk. Left monitor: a dark live monitoring dashboard
with four glowing line charts in cyan, violet, green and amber, a GPU
utilization bar near 100%. Center monitor: a before/after face-restoration
comparison, the left half blurry with ringing artifacts, the right half sharp
and clean with a green "aligned" checkmark. Right monitor: a black terminal
showing a CI pipeline where build, benchmark and verify stages are green and a
deploy step is running. On the wall, a translucent glass whiteboard with a
hand-drawn architecture sketch: three boxes labeled proxy, framework and engine
connected by arrows, and a dashed CI flow "tag → build → verify → deploy test".
A yellow sticky note reads "Heading home. Keep working after verify. Fix your
own mistakes. I'll check in the morning." To the right, a small server rack with
five glowing units: several Linux servers with blinking green status LEDs and a
Mac mini, one unit marked "2× GPU". On the desk: a mechanical keyboard with a
cyan underglow, an iPhone showing a long stitched screenshot, a steaming coffee
mug, and a compact camera. Futuristic glassmorphism style, dark navy background
with a subtle cyan grid, soft volumetric glow, neon accents in cyan (#22d3ee)
and violet (#7c3aed), clean vector-like shading, high detail, 16:9, 4k.
ChatGPT render: photoreal night office, engineer seen from behind at three monitors.
ChatGPT. Photoreal. Text is nearly perfect. Added a skyline, slogans, textbooks and dashboard metrics nobody asked for.
Gemini render: flat vector illustration, engineer seen from behind at three monitors.
Gemini. Flat vector. Nothing added, nothing missing. Dashboard labels came out garbled.
ChatGPT Gemini
Style Photoreal Flat illustration
Text Exact Mostly exact, small labels garbled
Extras Many None

ChatGPT treats a prompt as a brief to improve on. Gemini builds what you specified.

What is in the prompt

Almost none of it is decoration. The agent picked each detail from something we had actually done.

  • Seen from behind. It did not know my face and would not invent one.
  • Three monitors. A live monitoring dashboard, a face-restoration before/after, a CI terminal. The three things we spent the most time on.
  • “aligned” checkmark. A real bug: restored faces came back misaligned. We fixed it together.
  • proxy → framework → engine. The actual shape of our inference service, with the internal names removed.
  • tag → build → verify → deploy test. Our release flow. A git tag kicks it off; nothing goes to production without a manual step.
  • The sticky note. A message I left the agent one evening, word for word. It kept working overnight.
  • “2× GPU”. I keep asking it to use both cards. It remembered.
  • Long stitched screenshot. I send it phone screenshots stitched together so it can read a whole chat at once.
  • Compact camera. From the expo photos elsewhere on this site.
  • Cyan and violet, glass, grid. This site’s default theme.
  • Hoodie and headphones. Invented. The only guess in the whole prompt.

Neither picture is what I look like. Both are what my work looks like from inside an AI agent.