Near Leg,
Red Leg.
Far Leg, Blue.

I couldn't draw it. I could only tell you it was wrong.

That's the honest starting position. No art school, no animation background, no vocabulary for any of this. What I had was a game that needed characters to walk across a phone screen, an image generator that produced genuinely beautiful frames, and an unshakeable sense that the result looked bad.

My first note in the whole saga isn't a technical critique. It's someone squinting at a loop and describing a feeling:

Twenty minutes later I'd stopped being diplomatic:

Reading these back, I notice the thing I was best at was also the thing I'd have described as a weakness: I couldn't tell the agent how to fix it. I could only keep insisting, precisely and repeatedly, that it wasn't right yet. That turns out to be a real skill. It's just an expensive one to use as your only tool.

The first useful instinct was to build an instrument

Six minutes after "robotic and too clean," I asked for something that in hindsight shaped everything else. I couldn't see what was wrong at full speed, so I asked to slow it down and take it apart:

That request created the first of the little local review pages — play, pause, step, toggle, mirror — that every later decision got made on. I didn't think of it as tooling at the time. I just wanted to stop guessing.

It paid off immediately. Within half an hour I was no longer saying "it looks robotic," I was saying which frames were missing:

Same person, same lack of training, forty minutes apart. The difference is entirely the instrument.

The grind, and the wall at the end of it

Knowing the mechanism didn't fix it. It just made me a more precise complainer. The night of the 18th is the low point, and it's worth showing at full length, because the realization doesn't make sense without the pain that produced it.

Then this one, which is my favourite thing I said all week — because it's a person with no art training doing forensic occlusion analysis on a hip flap:

Eight minutes after that, I gave up on reviewing frames:

This is the "this isn't working and isn't scalable" moment, and I want to be clear that it wasn't a flash of insight. It was exhaustion. I had spent two evenings and written sixteen separate defect reports, by hand, on individual frames, and every fix broke something adjacent. The realization was less "I've figured it out" and more "I refuse to do this one more time."

Designing for the reviewer, not the viewer

With a master in place I started making art decisions for a reason I don't think a trained artist would have reached for first: making the animation easier to check.

That's a style decision justified entirely by diagnosability. Chunkier thighs weren't prettier — they gave the overlap at the hip enough surface to read at phone size, for me and for any automated check. The same session added the boot fix, which came from pure observation rather than anatomy knowledge:

And a bit later, the trick that finally made F4 read correctly — which is a genuine illustration technique I arrived at by staring at the problem long enough:

I also cut the model down. The master had ten phases; watching them play, two of them weren't earning their place:

That's why the shipped cycle is eight frames and not ten.

Then the arms, which I'd been quietly ignoring

The arm problem had actually surfaced a full day earlier and I'd let it slide:

Once the legs were solved by a model, the fix for arms was obvious — do the same thing again:

Revisions five and six were almost entirely arm surgery. The failures were consistent and strange, and I logged them one by one:

• F2/F3Frozen then jumping. F2 and F3 should have more red arm movement forward. The halfway point from F2 and F4 on the red arm.
• F4Spine arm. F4 has an extra arm growing out from the character's back. The far arm should be fully hidden behind the torso and cape — the generator preferred inventing a third shoulder to showing one visible arm.
• F5Growing limb. F5 seems to have increased the length of the red arm noticeably.
• R1Both again. R1 still has the arm coming out of the character's spine. red arm length also appears off.

Then I stopped:

I'm glad that's on the record, because "good enough, now how do we generalize it" is the correct instinct and I nearly didn't have it. Perfect was available for another six hours. Good enough was available immediately, and good enough is what everything else could be measured against.

"Good enough" is a threshold, not a compromise

I should be straight about what you're looking at on this page. This art isn't gorgeous. There are rough edges, a few bumps, a weird pixel or two if you step through slowly enough. That isn't an apology — it's the target, hit deliberately.

This is a prototype. Art is one of the cheapest things in the project to improve or replace later, and all of it eventually will be. What can't be faked at this stage is whether the world coheres — whether moving through it feels like the pieces connect, like you're absorbing a place rather than clicking past assets. That feeling is the real tell, and graphical detail is only one input to it. I'd argue it's among the lowest rungs on the ladder.

That's the whole standard, and it explains the shape of everything else in this journal. A walk cycle that reads wrong is a defect, because an animal I can't read pulls me straight out of the world. A slightly stiff elbow on F5 is a rough edge, and rough edges are affordable — they cost nothing that matters at this phase.

So every tick on that chart is me drawing that line on one specific frame. Not "is this beautiful?" but "does this break the spell?" Sixteen times over two evenings the answer was yes, which is why the grind happened. And when I finally wrote it's not perfect but it's good enough, I wasn't lowering the bar. I was saying the frame had cleared it — and that further polish was no longer the highest-value thing I could spend an evening on.

If you step through the sprites above and spot something imperfect, you're probably right. The question I'd ask back is whether you noticed it while it was playing, or only once you stopped it and went looking.

The second realization: write down how to catch it

One minute after "good enough," I said the thing that turned a working cycle into a working process. If the first realization was "build a model," this is the one I'd argue mattered more:

That's the moment the generation prompt and the rejection criteria became two different documents. Everything I'd learned the hard way got written down as things to look for, per frame, in advance — not as things to hope for.

It proved itself twenty minutes later on the next character. I caught a failure that passed every automated check, and rather than just asking for a fix, I asked for the check to change:

That single rule — compare three frames, not two, and reject a pair where the limbs effectively repeat and the next frame has to catch up — is now the hardest gate in the pipeline. A different file is not proof of advancement. Cloth moving isn't progress in a gait.

Then dozens of characters — rangers, clerics, mages, royals, workers, children, merchants — went through the same two-reference generation and the same triplet review. Twelve hours later:

What a non-artist can actually bring

I want to be careful not to tell this as "I was clever." I wasn't. I was stubborn, and I did the expensive thing for two full evenings before I did the cheap thing. The lesson is in the shape of the curve, not in any single insight.

But I've stopped thinking the lack of art training was the handicap. What I had instead turned out to matter more:

1. Taste you can't yet explain still counts.

   They walk like a mannequin is not a useful instruction, but it is correct, and refusing to accept output I didn't believe is what kept the project honest long enough to find the real fix.

2. Build the instrument before the artifact.

   Asking for speed control and per-frame toggles on night one is the reason every later judgement was specific. You can't fix what you can only see at twelve frames a second.

3. Pain is data. Count it.

   Sixteen hand-written defect reports across forty-eight hours wasn't a sign I should try harder. It was a measurement telling me the approach didn't scale. I ignored that reading for about a day and a half longer than I should have.

4. Get a second opinion that has no context.

   An agent three nights deep in your project carries its own habits and precedents, and gets stuck in them for the same reasons people do. When I needed animation theory rather than another revision, I asked somewhere that had never seen my bear. Knowing what to ask for was mine; the words were not.

5. Make the reference ugly on purpose.

   A beautiful reference gets copied for its beauty; a diagram gets copied for its geometry. Two references with explicitly separate jobs, and say in the prompt which one wins on pose.

6. Build a translation key, not just a reference.

   I could always see the error; I couldn't name it in terms the agent shared. Permanent identities — near hind, far fore — with a permanent visual cue carried into the finished art give both sides the same four objects to point at. That's what turns "this looks wrong" into "this limb moved wrong."

7. The audit is the deliverable.

   The generation prompt and the rejection criteria are different documents. Writing the second one is what turned one good warrior into dozens of good characters and, a night later, into a bear.

8. Ship "good enough" and generalize.

   "Good enough" is a threshold, not a shrug: does this break the player's connection to the world, or is it just imperfect? Revision six was imperfect and cleared the bar, which made it worth more as a baseline than a seventh revision would have been as a polish pass.

Four nights, three hundred-odd frames, and the thing that actually fixed it was a grey blob with one red leg and one blue leg. Sep 16–19, 2026