Releases·Tool Release
Show HN: Pac-Bench – How well can models one-shot a Pac-Man game?
Pac-Bench, shared via Show HN, benchmarks how well model+harness combinations can create a Pac-Man game from a single prompt — 'Create a Pac-Man game in a single HTML page'. Each model gets one shot with no follow-up prompts or fixes, testing genuine single-prompt coding ability. It matters to developers comparing frontier models for unassisted code generation.
127 engagement1 sourceMon, Sep 28, 2026, 10:43 PM
AI Overview
- Pac-Bench (jonclegg.github.io) tests how well Harness+models one-shot 'Create a Pac-Man game in a single HTML page' with no follow-up prompts or fixes.
Related
Social·Community
I made 20 one-shot HTML5 game prompts for testing local coding models
A Reddit user published a set of 20 one-shot game prompts for testing local coding models, each requiring a complete game in a single index.html with no external libraries, assets, APIs, or internet access. The prompts vary in complexity, covering enemy AI, procedural generation, upgrade systems, bosses, shops, and physics. The author invites others to run the same prompts across different models and harnesses to compare results.
A Reddit user published a set of 20 one-shot game prompts for testing local coding models, each requiring a complete game in a single index.html with no external libraries, assets, APIs, or internet access. The prompts vary in complexity, covering enemy AI, procedural generation, upgrade systems, bosses, shops, and physics. The author invites others to run the same prompts across different models and harnesses to compare results.
17 engagement1 sourceredditFri, Oct 2, 2026, 02:56 PM
Developments·Benchmark
Neon Ladder: a playtest-graded benchmark for local LLM stacks — your config is the subject, a working game is the grade
A developer released Neon Ladder, a public benchmark that grades local LLM stacks by having a coding agent build a 10-file HTML5 canvas game from a fixed contract. Three gates score the result: static code checks, a two-minute headless browser soak test, and human playtesting. The author reports it catches failure modes that llama-bench and PPL miss, and is soliciting community numbers.
A developer released Neon Ladder, a public benchmark that grades local LLM stacks by having a coding agent build a 10-file HTML5 canvas game from a fixed contract. Three gates score the result: static code checks, a two-minute headless browser soak test, and human playtesting. The author reports it catches failure modes that llama-bench and PPL miss, and is soliciting community numbers.
3 engagement1 sourceredditThu, Sep 3, 2026, 04:36 PM
Social·Community
Community proposes DOOM as a one-shot benchmark for AI agents
A Reddit user suggests using DOOM as a benchmark to test AI models and harness setups, arguing it requires planning, architecture, programming, visual judgment, debugging, tool use, scope control, and knowing when a task is finished. The proposal highlights gaps in traditional benchmarks and emphasizes fairness, suggesting a browser-based game to prevent cheating.
A Reddit user suggests using DOOM as a benchmark to test AI models and harness setups, arguing it requires planning, architecture, programming, visual judgment, debugging, tool use, scope control, and knowing when a task is finished. The proposal highlights gaps in traditional benchmarks and emphasizes fairness, suggesting a browser-based game to prevent cheating.
13 engagement1 sourceredditSat, Aug 1, 2026, 01:32 AM
Developments·Benchmark
Show HN: Slime Soccer benchmark reveals flagship models don't always dominate
A developer created a benchmark where AI models play Slime Soccer, a simple web game, to compete directly and visually. The results showed that flagship models like GPT 5.6 Sol performed poorly, surprising the creator.
A developer created a benchmark where AI models play Slime Soccer, a simple web game, to compete directly and visually. The results showed that flagship models like GPT 5.6 Sol performed poorly, surprising the creator.
9 engagement1 sourcehackernewsTue, Jul 14, 2026, 02:24 PM
Products·building
Open-source harness that lets Codex supervise Claude coding agents
A developer is building an open-source harness that pairs Claude with Codex, where Codex acts as a 'babysitter' overseeing Claude's output. It aims to remove the human from the loop of supervising AI coding agents. A test game was generated in one shot from a single prompt.
A developer is building an open-source harness that pairs Claude with Codex, where Codex acts as a 'babysitter' overseeing Claude's output. It aims to remove the human from the loop of supervising AI coding agents. A test game was generated in one shot from a single prompt.
3 engagement1 sourceredditTue, Sep 8, 2026, 12:06 AM