Bench Vibe Coding Apps (Game Loop Grade)

Vibe coding apps ship demo GIFs of a paragraph turning into a game. The GIF ends before hour two, which is where every weekend jam actually gets decided. This post benches vibe coding apps on the axes that matter for a playable loop — five pass/fail rows first, then a head-to-head between chat-only surfaces, IDE-embedded assistants, and WizardGenie, plus how AI Image Gen hooks in without asking the agent to rewrite the scene. Coding model names match the Sorceress CODING_MODELS catalog in src/app/_home-v2/_data/tools.ts, verified August 3, 2026. Phaser v4.2.1 "Giedi" was live-checked the same day against the official GitHub release notes.

What vibe coding apps means for a jam

DataForSEO lists vibe coding apps at 1,300 monthly searches with KD 23 in research-supplement.md, verified August 3, 2026. The plural matters: the reader is not shopping for a single CLI they can install once and forget. They want a shortlist of concrete applications — browser tabs, desktop installers, IDE extensions — that they can open on Friday and still trust on Sunday. The sibling post that graded whole vibe coding platforms covered the surface level; this article drops one level and benches the apps that actually run on your machine.

Wikipedia's vibe coding entry (checked August 3, 2026) traces the term to Andrej Karpathy on February 2, 2025 and marks it Collins English Dictionary's Word of the Year for 2025. Martin Fowler separates pure "forget the code exists" vibe coding from agentic programming where structure still matters. A jam game sits firmly in the second bucket. You may not hand-type every line, but you absolutely care whether gravity, hitboxes, restart rules, and named constants survive the next prompt.

So for a weekend build, the query collapses to five demands. Any vibe coding app worth putting on the shortlist must let you keep one project folder warm across sessions, show a live preview beside the chat, accept a scoped feel fix without regenerating the world, swap sprites and SFX as file drops, and split expensive reasoning from cheap typing. Apps that fail any one of those demands are demotion candidates for a game. The bench in the next section formalizes that into a fixed scorecard.

Fixed bench: five pass/fail rows

Ignore the landing page trailer. Score five rows on paper before you commit a weekend to any candidate. MDN's overview of web games still centers on a running loop, input handling, and asset intake — a fair bench for vibe coding apps has to center on the same surfaces, not on the app's marketing copy.

Check	Pass looks like	Fail looks like
Playable slice	Move, collide, score, restart in one screen	Pretty menu with no loop
Single-variable fix	One constant changes; rest stays	Whole scene regenerates
Asset add	One PNG + one SFX drop in cleanly	"Rebuild the game with art"
Reopen continuity	Day-two folder still runs	Disposable sandbox evaporates
Cost split	Frontier plans, cheap model types	Frontier model on every gravity tweak

Two housekeeping rules make the bench honest. First, pin a runtime before you scaffold. As of August 3, 2026, the stable Phaser release on phaser.io/download/stable is Phaser v4.2.1 "Giedi" (released July 9, 2026 per the GitHub release notes). Naming the version in every scaffold prompt keeps physics behavior identical across every app you grade.

Second, run one prompt against every candidate. Vary the prompt to flatter one app and the bench turns into marketing. Paste this into each candidate verbatim:

Build a one-screen browser action prototype on Phaser v4.2.1. Side view. Player moves and collides with platforms. Collect three pickups to raise score. Touching an enemy restarts the run. Colored rectangles only. No inventory, no dialogue. Expose moveSpeed and enemySpeed as named constants. Do not invent extra systems.

Then demand a single-variable fix: "Enemy speed 90 feels unfair on the first screen; lower to 70 and keep player speed fixed." Any app that rewrites the world for that request gets a red mark on row two. The rest of the sections walk each app class through the same five rows.

Chat-only apps vs playable-loop apps

Chat-only vibe coding apps are the loudest category. One input, one output, and a landing page that shows a full game appearing out of a paragraph. They earn narrow wins and predictable failures on the bench.

Wins: throwaway UI demos, one-shot experiments, and non-game side quests. If you need a landing page, a CRUD mock, a pitch deck helper, or a sixty-second silly minigame you will never open again, a chat-only app can be faster than any engine. That is a real use case; do not throw the whole category out. Marketing pages, jam trailer sites, and README landing pages are all fine.

Failures: everything the bench actually measures. Row-one slice often passes on the first prompt and then collapses on the second, because "add a second loop" is heard as "start over." Row two fails constantly, because "lower enemySpeed to 70" is heard as "regenerate the scene with slower enemies." Row three fails when you drop a sprite in and the app rewrites the collision code around a new atlas. Row four is the killer: without a durable project folder, day two is a fresh disposable sandbox with none of yesterday's tuned constants. Row five may pass on unit price but usually fails on architecture, because the app only exposes one model and no planner/executor slot.

Playable-loop apps take the opposite trade. They keep a live preview beside the chat, hold a project folder across sessions, and expose named constants so a scoped fix stays scoped. On the bench, they typically pass rows one, two, and four cleanly, and pass row three when the asset intake is real. They win the weekend even when their chat window looks less flashy than a chat-only app's landing page. Grade the loop, not the chat.

Where WizardGenie earns a top-row pass

WizardGenie is the Sorceress AI-native game engine: describe the game, watch it run, iterate in real time. It ships as a Windows desktop installer and as a web entry at /wizard-genie/app. Both paths share the same project format, so a bench run on either path counts for the same grade, and the dual path is why WizardGenie earns a top-row pass on reopen continuity without asking you to bring your own filesystem discipline.

Row by row on this bench: the playable slice passes because the product is built around a live preview, not a code dump. The single-variable fix passes because sacred constants (moveSpeed, enemySpeed, jump height, restart rule) stay in the project instead of vaporizing when you mention art. Asset add passes because you can hand it a PNG plus a texture-swap instruction and keep the loop running. Reopen continuity passes on desktop (the folder is real) and passes on web when the project is saved to your account. Cost split passes because WizardGenie exposes a Planner and an Executor slot, so a frontier model plans and a cheap model types.

The Sorceress tools guide maps the non-coding surfaces that pair with the agent on the same account and project habit: image, audio, and 3D tools that a WizardGenie prompt can point at without asking the agent to invent anything new. That pairing is what turns a graded app into a graded pipeline. IDE-embedded assistants sit inside your existing editor and pass reopen continuity trivially, but they usually lose row one (no play surface) and row three (asset add is a manual copy), and they force alt-tab friction on every micro fix.

Cost split without a frontier executor

Cost split is the row that separates vibe coding apps you can afford on Sunday afternoon from vibe coding apps that quietly turn a weekend into a bill. The economic logic is simple: put frontier reasoning on the Planner side for architecture, hard bugs, and "why does this collision feel wrong," then put a genuinely cheap model on the Executor side for the twenty gravity tweaks that dominate day two.

Verified against Sorceress CODING_MODELS on August 3, 2026:

• Strong planners: Claude Opus 4.7, GPT-5.5, Gemini 3.1 Pro, Grok 4.2 (Claude Sonnet 4.6 when you want a faster planner tier).
• Practical cheap executors: DeepSeek V4 Pro, Kimi K2.5, MiniMax M2.7.

Never park a frontier-priced model on the typing pass. That pairing wipes the cost advantage that makes dual-agent work worth doing, and it trains you to fear iteration. The rough shape you want is Opus or GPT-5.5 thinking while DeepSeek or Kimi types the diff. A clean prompt pattern for row-two safety:

Planner: propose the smallest change set that lowers enemySpeed from 90 to 70 without touching playerSpeed, jump, or collision shapes. List files and constants only.
Executor: apply only that change set. Do not refactor unrelated systems.

If the executor rewrites the scene graph anyway, that is an app failure, not a "vibes" failure. Row two of the bench is designed to catch exactly that pattern. The higher an app's row-two pass rate under a scoped planner/executor prompt, the more you can trust it with the feel-tuning that dominates jam hours three through ten. Apps that hide the model picker or lock you to a single tier fail this row on architecture, not on price.

Art hooks that keep constants intact

The final honest test for vibe coding apps on a game bench is what happens when art enters the project. Row three (asset add) fails on most chat-only apps because the request "swap the player sprite for this PNG" is heard as "regenerate the scene with new art." A playable-loop app treats the PNG as a file, updates a texture reference, and leaves every named constant on the game loop untouched.

The Sorceress art hook that keeps the bench honest is AI Image Gen for a reference-locked sprite plus a matte cut, handed back to WizardGenie as a texture swap instruction. The prompt on the agent side is a scoped diff, not a regeneration:

Player texture is player.png at 32x32. Swap the current placeholder to that file.
Do not modify collision shape, moveSpeed, enemySpeed, jump, or restart rules.

Under that pattern, art becomes a file drop, not a rewrite. Audio follows the same pattern: a WAV goes into the assets folder, a scoped instruction wires it to a specific event, and no other system moves. Chat-only apps often break this because they treat the entire game as a single generated artifact; a scoped file swap is not in their vocabulary. Playable-loop apps that treat the project as a folder pass row three every time.

Grade the app on how it responds to "add art without changing feel." That single request is the fastest reveal of whether the app understands games as durable projects or as one-shot generations. Every other row of the bench measures the same underlying property from a different angle: durability of the project across time, prompts, and files.

Verdict for a weekend build with vibe coding apps

Grades collapse to picks. Here is how the bench sorts vibe coding apps for a Friday-to-Sunday build:

Need	Graded pick	Why
Playable jam slice with live preview	WizardGenie (web or desktop)	Chat and play stay in one loop
Hard architecture or weird bug	Frontier planner in WizardGenie	Opus / GPT-5.5 / Gemini 3.1 Pro / Grok 4.2
Twenty feel tweaks	Cheap executor slot	DeepSeek V4 Pro, Kimi K2.5, or MiniMax M2.7
Roster or prop art	AI Image Gen → texture swap	No scene rewrite
Static marketing page only	Generic chat-only app	Games are not the bottleneck
Existing multi-thousand-line codebase	IDE-embedded app	Continuity beats live preview at that scale
Day-two continuity for jam projects	Desktop WizardGenie	Real folder beats disposable sandbox

Bottom line for the SERP query: the vibe coding apps worth putting on a shortlist for a weekend game build are the ones that pass the five-row bench, split planner and executor honestly, and let AI Image Gen swap textures without rebuilding the world. Grade WizardGenie first when you need a playable agent path. Keep chat-only apps for non-game side quests. Reserve IDE-embedded apps for the codebases where the play surface is already yours to open. Re-verify model names and Phaser pins on the day you ship — today's check was August 3, 2026.

When you are ready to run the loop, open WizardGenie, pin one screen, and refuse any prompt that regenerates the world to change a single number.