The AI Loop Method: every week, an AI designs my drawing lesson. I follow it in colored pencil — a celebrity portrait, start to finish. Then the finished drawing goes back to the AI for a graded verdict: HIT or MISS. The next lesson targets exactly what the critique found.
Running publicly since May 2026 at Pencil & Prompt. Every verdict below happened on camera.
Weeks 1–4 were pre-method groundwork. The loop record starts at week 5, newest first. Binary HIT/MISS verdicts were formalized in week 9. Click any portrait to see it full size.
| Week | Subject | Target set by the previous critique | Verdict | What surfaced next |
|---|---|---|---|---|
| 11 | Jude Bellingham | Turn the flat face sculpted — peak highlights + core shadows | HIT — with one catch | Orbital shadows overshot into hollow; highlights read scratchy |
| 10 | Taylor Swift | Warm, controlled mid-value shadow + warm lip base | HIT | Face reads flat — no true peak highlight |
| 9 | Lamine Yamal | Turn the away-side plane of the face | HIT — with one catch | Plane turned, but went grey and heavy doing it |
| 8 | Cristiano Ronaldo | Turn the shadow side as convincingly as the lit side | Partial — pre-verdict era | Patchy skin, three weeks running |
| 7 | Lionel Messi | Warm-but-controlled darks + beard density gradient | Two threads landed — pre-verdict era | Shadow side flattening |
| 6 | Jalen Brunson | Cool the darks + unify the skin | Partial — pre-verdict era | Overcorrected straight into grey |
| 5 | Kylian Mbappé | Baseline — the record starts here | Baseline | Shadows too heavy · skin patchy · stubble a flat mass |
Single verdicts are the easy part. The useful record is what happens when the same flaw survives several corrections. Three threads so far:
The best form modelling of the series: the face finally reads lit and three-dimensional, and the shadows stayed warm this time, so week 9's grey lesson held. The catch is the oldest habit on this page — told to add core shadow, I drove the eye sockets past the reference into hollow. Five weeks now of overshooting the moment a shadow gets named.
The patchy-skin fight I'd been losing since week 5 is finally over — the AI ruled the thread closed, the most unified skin of the series. Then it immediately found the next thing: the face is unified now, but it reads flat.
The away side of the face finally turns — the thing I'd missed two weeks running. It also went grey and heavy doing it. Hit the target, paid a price. Best curly hair I've ever rendered, though.
Most ambitious angle yet — an upward three-quarter — and the likeness survived it. The skin didn't: patchiness three weeks running, and the AI called it out as the standing thread.
Two fixes landed at once: the beard finally grows out of the skin instead of sitting on it, and the darks hit the warm middle after two weeks of swinging. Then the AI found the shadow side flattening. There's always a next thing — that's the point.
Fixed the too-hot darks by overshooting straight into grey. The dial has two directions and I found both. Skin did come out more unified than week 5 — real progress hiding inside a miss.
The first portrait ever put in front of the AI for grading. Its first fix list was long: shadows too heavy, skin patchy, stubble a flat dark mass. Everything above this card is the story of working through that list.
Weeks 1–4 and early one-off portraits, drawn before the grading began. Where the record above started from. Click to view full size.
Want to run the loop yourself? Every episode's lesson PDF is free. If you want to go further with your own reference photo, the Custom Color Map is the same AI analysis applied to a photo you choose — the exact colors and sequence for your drawing, your pencil brand.
Everything lives at pencilprompt.gumroad.com.
After 11 documented weeks: it can direct and grade practice — remarkably well. It names the specific flaw I keep avoiding and holds me to one target a week. What it can't do is replace the hours; week 9's catch was pure hand execution, and so was week 11's. The record above, misses and overcorrections included, is the honest answer so far.
It compares the finished drawing against the reference photo and the lesson's stated goal, then rules on one specific target — likeness, values, form structure — in plain language. One binary verdict per week: HIT or MISS. Never a numeric score, because a "7/10" teaches nothing; "the away-side plane finally turns, but it went grey" does.
Two of the three long threads are closed. Patchy skin, flagged in week 5, was ruled closed in week 10. Flat, unsculpted faces — flagged in weeks 6, 7 and 10 — closed in week 11. What's left isn't a missing skill at all: it's a habit. Five separate weeks, the AI named a shadow and the correction went straight past the target. That's the whole current fight.
On this record, almost never because there wasn't enough shadow. The cause was lit planes — forehead, nose bridge, cheekbone — never reaching a true peak highlight, so the face read as evenly filled instead of struck by a light source. Adding more dark made it muddier, not rounder. What finally closed it in week 11 was a hard value jump on those three planes plus gentle core shadow — which is the opposite instinct to the one most of us have.
Because the shadows were built with neutral or cool darks. A real shadow still receives warm bounced light, so a dead-grey shadow reads as an absence of light rather than a form turning away. Glazing warmth into the shadow masses fixed it here. It showed up in weeks 6 and 9 — both times as an overcorrection to being told the darks were too hot — and has held clear in weeks 7, 10 and 11.
Those tools grade an image once. They never find out whether the advice worked, because almost nobody returns with the same drawing corrected — so what they hold is advice issued, at scale. This page is the other shape: one person, one medium, every correction actually executed by hand and then re-graded the following week. What's published here is advice tested — including the weeks where following it made the drawing worse. It's a much smaller sample and a much slower one. It's also the part a tool can't produce.