On 13 October I sent the agents a long list of changes from my kids' playtests. Most of it was about getting lost. One line was about the world itself. The next quest I was given sent me straight back into the hedge maze: "I didn't want to do the hedge maze again, I wanted charred wood, or castle... users should go through the various mazes / worlds rather than repeating."
Further down I tried to say what the game is: "We have a maze generator, but that whole storyline about shifting mazes / shifting worlds etc needs onboarding... The core premise of the game here is mazes around the core story arc." The list ended: "Let's start work on creating the other worlds / mazes (charwood) and also more of the characters / monsters etc... Let's cook!!!"
When the root agent asked where the main line should go after Tilly's remedy, I picked the Charwood. It was built in an afternoon, rebuilt that evening, then tuned twice more once we had played it.
A wood from the lore
The lore already had the Charwood: an old oak and birch forest with charcoal clamps, where Tilly's father Jory burns charcoal. The agent read my "charred wood" as a working burners' wood rather than a burnt waste, friendlier for children and truer to the lore.
It stayed generated, because the premise wants it. Walk to the end of North Lane and the game makes a fresh wood, unless a quest wood is waiting. Go back the same day before finding its heart and you get the same one. The first version borrowed the hedge garden's machinery, which is why it was playable within hours.
The generated hedge garden, whose square-grid maze the first Charwood borrowed.
It got wolves, wild boar, bramble-backs and Old Greymane, the pack leader in a paler coat. I had called our blacksmith "a rubbish meshy model", so Bram and three other townsfolk were remade too.
Under the hood Meshy only rigs two-legged models. A new script fits a free CC0 Quaternius wolf or fox skeleton to each static Meshy model, copies the skin weights by nearest face and keeps that skeleton's own walk, gallop, attack and death clips.
"Too much like the hedge maze"
That afternoon I asked: "Can you give me a simple screenshot of part of Charwood quest, so I can see how it looks". Two notes went straight back: "The way the trees disappear is really weird", and "I don't like the repeated trees all being the same. Let's add in at least 4 or 5 other trees, and some other bush types etc, let's make it more organic and varied."
A minute later I put my finger on the real problem: "It feels too much like the hedge maze, let's make it feel different. It feels very sparse? Is this on purpose? I'm wondering about more 'mood' here, but maybe the sky / haze will help with this."
The root agent's answer was honest. It felt like the hedge maze because it was the hedge maze: the same square grid, with trees standing in for hedges. Haze would help, but it would not fix the layout.
The first Charwood, already with nine tree kinds. The grid shows through.
At eye height: a straight corridor between rows of bushes.
Under the hood The first pass of the wood drew 12.2 million triangles in 240 draw calls on High. Lower-detail tree models for distance brought that down to 1.58 million in 88.
The root agent offered a fresh piece for a proper woodland. I said: "no, let's do that Charwood build."
A woodland, not a maze
The rebuild started with the layout. A new woodland generator grows irregular clearings and joins them with winding rides, spurs to dead ends and the odd loop. Trees are planted by distance from the path: ferns at the edge, then undergrowth too thick to see through, then trunks and broad oaks. A canopy you never see casts dappled shade. There is a low warm sun, drifting ash, smouldering burners' clamps and distant wolves.
Under the hood The woodland generator writes the same layout format as the maze, so the solvability checks, the save record and the quest's way in did not change. Woods already saved on the old grid still open, re-dressed as the new wood.
Playing it caught what the unit tests had not: no clamp ever appeared, and several seeds made the same wood. A clamp on a junction blocked the way, so the generator quietly skipped to the next seed. Clearings now open a small plaza round their centre.
The rebuilt wood that night. The pale streaks are sun shafts seen from above.
At eye height: layered thicket, a grey-amber sky and no straight lines.
By then I was on a train: "Don't upload now, I'm on a train." Then: "193Mb seems like a lot, is that properly ktx2'd and compressed etc". It was. The 193 MB were the raw Meshy masters, kept for rebuilds.
Under the hood What players download for the new models came to about 14 MB, with every texture in KTX2. The whole Charwood batch spent about $1 of a $50 OpenAI cap and 230 of 3,000 Meshy credits.
Both pieces were merged before midnight.
Playing it
The next morning I asked: "can I go to Charwood to check its look feel without doing all the quests?" Then the notes came. "In Charwood, all 5 wolves appear together at the same time, and I was done. Should these be placed more apart, maybe a pack of 2 and a pack of 3?" Worse: "The 3 wolves here aren't attacking me or coming towards me? They all turn as I move, looking at me, but they should run (faster than my run) and want to attack me."
The cause was a small mismatch: pack wolves were spaced as circles but collided as boxes, so they stood pressed together and every step was refused. Now they come in packs of three and two, and they outrun a sprint.
I also asked: "Let's not do any disappearing trees in Charwood, keep them always in display, I can always go first person to see better." The fade went.
Too dark to read
On 17 October I asked for a rethink: "It's too dark and it's really, really complicated to know which way you can actually walk... Maybe the grounds are different colour on the ways you can walk rather than the bushes." And: "there's a lot of just wandering around doing nothing until you find some wolves." The end confused us too: "we killed all the wolves, killed Old Greymane... and then we had to find the clearing. But we found the exit and went back to the village."
The agents took the hint about the ground: darker leaf mould under the thicket, paler beaten rides. The light went up, the haze thinned, and fireflies drift over the rides. Little monsters turn up along the way, and once the wolves are down any way home finishes the quest.
The same spot and seed: 13 October (left) and 17 October (right).
What I took from it
Reusing the hedge garden got the Charwood playable in an afternoon, but borrowed machinery brings its borrowed look. More trees would not have hidden a grid. The fix was the layout, and because it wrote the same format, nothing downstream noticed.
Mood and readability also pull against each other. On the Tuesday I asked for "more mood". By Saturday I wanted to see where I could walk.
But the wood was only one stop. Back from the train on the 13th, I had already asked for the rest: "After you've done the Charwood work, carry on creating the quest and flow and mazes / side quests / story arc."