A simple prompt to “build a pixel-art Pac-Man game in HTML” failed to produce a satisfactory result from Fable, even after several rounds of added instructions. The case is small, but it supports a clear judgment: the frontend capabilities of today's generative AI cannot be measured by whether it can produce some code.
What this is
This is the Pac-Man Benchmark posted on r/LocalLLaMA. The tester asked the model to create an HTML version of Pac-Man and specified a pixel-art style; even after adding more prompts, the results still fell short of expectations. The post did not disclose the full prompt, runtime environment, or model version, so it is better understood as an individual observation than a rigorous benchmark.
What the test actually probes is whether a model can understand the gameplay, interface layout, and visual style at the same time. A prompt—the user's instructions to the model—may look like only a few sentences, yet it directly affects whether the code works, whether it meets the requirements, and whether repeated generations remain consistent.
Industry view
We believe the value of a small test like this lies not in ranking models, but in reminding teams to establish finer-grained acceptance criteria. Frontend development usually involves interaction, visual fidelity, and browser compatibility. Looking only at whether the code runs can obscure the gap between code that works and code that is actually usable.
The risks are also clear: without public prompts, runtime environments, scoring rules, and records of repeated runs, results are easily skewed by chance. A model's poor performance in one test does not prove that it is inherently weak at frontend development. Conversely, a polished demo page does not mean the model can consistently complete real-world projects.
For product teams, a more reliable evaluation method is to fix the task, repeat the run, and score functional completion, visual similarity, code maintainability, and human modification time separately. Conclusions drawn this way are much closer to procurement and deployment decisions than those based on a single impressive demo.
Impact on regular people
For enterprise IT: When bringing AI coding tools into the organization, procurement evaluations should include first-pass success rates and rework time, rather than relying on demo performance alone.
For individual careers: Nontechnical roles can describe requirements in natural language, but complex pages still require human judgment about screenshots, interactions, and code quality. Writing clearer prompts does not eliminate the responsibility for acceptance testing.
For the consumer market: As more websites and small applications are generated rapidly by models, users will gradually shift from asking whether a page was handcrafted by a human to focusing on whether its functions are accurate and its experience is smooth.