Yes — and Aaron's framing is exactly right: no individual game needs the framework (the AI does the building); the framework's payoff is the shell/cross-platform/telemetry/consistency layer, which my benefits doc captured. The friction I hit was almost all "the skills are written for ports (mirror an existing game), so building net-new — especially a non-standard archetype — meant reverse-engineering from existing games." Here's the concrete list, ranked by how much churn each actually cost.

Real debugging loops (cost the most)

1. Mantle text-glyph recolor via color matrix renders invisible. I followed the skill's Mantle cheatsheet — setNodeColorMatrix(node, ZERO_MATRIX, offset) documented as per-instance recolor with "alpha passthrough" — to tint locked/hint letters. The glyphs came out invisible. Cost me a real bug hunt (isolate typed vs. locked, partial-type probes) before I switched to prewarming pre-colored glyph bundles + setNodeAsset (what clue-rush actually does for its 2 colors). Fix for the skill: warn that per-instance color-matrix recolor of text bundles is unreliable; document "prewarm one A–Z bundle set per color, swap assets" as the canonical text-recolor pattern.

2. Async text-node mount race → overlapping text. My first setClue/setHintLabel created a new text node per update; the async createTextBundle calls raced and left stale nodes mounted (garbled overlapping clue + LEVEL text). Fix: document the makeText-style pattern (one persistent node, setText to mutate) as the rule for any changing text, and explicitly warn against per-update node creation.

Missing archetype / reference coverage

3. No net-new scaffold or designated net-new reference. create-game says "follow port-lumos-game steps 2–7, skipping extraction," but there's no template and no canonical net-new reference called out — I copied boilerplate (package.json/tsconfig/game.ts/factory/run/text helper) by hand from clue-rush/word-bubbles/raindrops. Fix: a scaffold <id> generator (or a templates/game-skeleton/) + name the 1–2 canonical net-new references.

4. The untimed / endless / bespoke-HUD archetype is undocumented. Word Ladder is no-timer, no-level, own chrome. I had to discover by trial: streaming runner with no duration, an onSessionStart that awaits forever (ends on quit), enabled: { hud: false }, hudShowFirstSlot: false, streak.pipCount: 0, a no-op presentIntro. That's a legit game family (gym-machine games) with zero skill coverage. Fix: a "calm/endless game" recipe documenting exactly this.

Slowed me, resolved by reading source

5. createTextField for typing games — no getValue() (must track via onChange), lowercase defaults true, soft keyboard needs softKeyboard:{element}. It's the canonical text-input primitive (word-bubbles + both my games) but undocumented in the skill.

6. Text-bundle sizing contract — globalBBox is EM-ish (~0.7), not px, even with fontSizePx; scale = targetH / bbox[3]. I learned it from a stray raindrops comment. Plus the existence of makeText/placeTextInRect helpers isn't surfaced.

7. Streaming-session-end semantics — timed (duration + awaitTimeout) vs. fixed-count (loop then return) vs. endless (await-until-abort), and what outcome/finalize each yields. I inferred the endless pattern.

8. Coordinate contract — referenceUnit-normalized viewport (short edge = 100), pixelToUnitXY, viewport.subscribe relayout; games mix center-origin (word-bubbles) and top-left (raindrops). Took reading several games.

9. Dynamic-subtree teardown idiom — scope.child() + new Node({scope}) + scope.add(() => parent.removeChild(root)) + scope.close() for rebuild-per-trial. A leak footgun; learned from clue-rush.

Tooling & contracts

10. The verifier driver is click-only. For typing games I had to hand-write a keyboard driver and add a temporary debug hook to reach game state. Fix: add --type "WARM\n" / key-sequence support to the bundled driver.

11. Net-new payload contract is fuzzy — "ship the same shape as the closest catalog game," but shapes vary wildly (word-bubbles flat vs. the Base.js envelope with stat). I guessed a flat object. Fix: state the minimum contract finalize expects and whether stat matters for net-new.

12. End-of-session UX is undefined — this is literally what caused the Crosswise "restart" rabbit-hole: it's unclear whether the shell shows results / play-again at session end (the results overlay exists but isn't mounted by default; the harness just goes idle). Documenting the end-of-round contract would've saved that whole investigation.

Net: the framework itself held up well — everything above I could solve by reading an existing game. The improvement is closing the gap between "port a Flash game" (well-covered) and "build net-new, possibly non-standard" (reverse-engineering).

If useful, I'm happy to turn this into concrete PRs against the skills — e.g. add a "net-new + calm/endless recipe" section and the Mantle-text gotchas to create-game, a makeText note + createTextField doc to the rendering cheatsheet, and the --type flag to the verifier driver. Want me to draft those?