Test prompts for new and local LLMs
62 hand-picked prompts that separate strong models from weak ones, from 11 sources: 36 build tests (simulations, games, apps, agent tasks) and 26 quick checks with verified answers. Every entry says what it tests, how to grade it, where small models usually fail, and exactly where it comes from.
62 of 62 prompts
Coupled subsystems (cells, three resource pools, brood cycle, bee roles), a fixed 10 ticks/s logic loop decoupled from rendering, a priority order for features, and a …
3D rendering plus vehicle physics (inertia, drift, surface grip), AI riders that follow the track and make mistakes, lap timing and a debug overlay, all in one HTML …
Pointer lock and mouse look, hit detection by raycast or projectiles, enemy AI with patrol, chase and line of sight, two distinct weapons, menus, and a start-up …
Whether the model can deliver a compilable, playable C++ game in one file: it has to pick and state a graphics approach, then build physics, tricks, grinding, combos …
Building a real 3D game in Python with the easy library (Ursina) forbidden, so the model has to choose and use another stack correctly.
Game-design depth: four towers with distinct roles, enemy types with resistances, an economy with wave scaling, a map editor and an automatic balancing assessment …
Data modelling for item flow (miners, belts, machines, storage, power) and an automatic bottleneck detector, the kind of logic small models usually fake.
Three trophic levels with energy, ageing, reproduction and death, heritable traits that mutate, environmental pressure and live population charts.
A road network with intersections, vehicles routing to destinations, signals or priority rules, emergent jams, and user changes whose effects must show up live in the …
Simplified but internally consistent orbital mechanics with thrust manoeuvres, fuel, trajectory prediction and a mission goal.
Coding-agent discipline in a real folder: explore first, follow existing file formats, no downloads, no moving or deleting files, validate the output and report …
Turning business rules into an importable n8n workflow with a classifier, branching, a review label for paid offers, placeholder credentials and a dry-run switch that …
A large spec with an explicit acceptance list: a hierarchical robot rig, tripod gait, arm and claw, sensors, a drone, mission presets, telemetry, and a …
A real application rather than a demo: layer compositing with per-layer canvases, many tools, filters on the active layer only, undo/redo and export.
A short, ambitious prompt: seamless zoom from space to street level, country hover with statistics, clouds, air traffic and day/night. It tests honest handling of data …
Open-ended game creation that must avoid protected names and designs while still delivering exploration, catching, turn-based fights, levels, evolutions, items and …
A 54 KB master prompt for a coding agent with binding safety limits, strategy research, backtesting, an external audit, paper-trading deployment and an explicitly …
An autonomous website-builder prompt: from one product photo to a polished scroll-driven landing page with a fixed section order, OKLCH palette, copy rules, browser QA …
A compact spec with a clear success criterion: first-person movement, terrain generation, mining and placing blocks, several block types, collision and gravity.
A playable third-person fantasy prototype with combat, enemy AI, character stats, abilities, experience and level-up.
A research-plus-production task: about one minute of 16:9 video on Micron's latest quarterly report with figures, outlook, Hyperframes animation, a German ElevenLabs …
The ARES-6 robot again, now with hard physics constraints: limbs must not intersect each other or the body, the robot must not walk through objects, feet need …
Exact requirement recovery (20 balls, given colours, numbers, spin), hand-written collision detection and response against rotating walls, ball-ball collisions, …
A one-line prompt (from xAI's Grok 3 launch demo) that hides real orbital mechanics: transfer time, waiting for the return window, scale and 3D inclination.
Seventeen precise, checkable requirements in Chinese: a canvas ASCII Mandelbrot set, 0.5 % zoom per frame for 200 frames then loop, characters from Li Bai's poem 静夜思 …
A precise visual spec in Chinese: the Sun and nine planets in order with relative sizes, four given colours per body, labels, 1 s = 10 Earth days, an offscreen canvas …
A 2D particle fluid with hand-written collisions against a rotating container: leak-proof segment joints, restitution, particle repulsion and a coordinate transform …
A long, number-heavy three.js spec: 100 colliding firecrackers, timed fuses, shock waves, fire-radius ignition with random delays, particles, dynamic lights and three …
Modelling and materials (a real Erlenmeyer flask, physical glass, liquid that fits the shape) plus a foam particle system with decaying pressure, squash on contact, …
Ball-ball collision response with mass, damping and wall bounces for 50+ objects, delta-time animation and some optimisation, inside a fixed HTML boilerplate.
State management: wrap-around movement, collision detection between moving objects, an explosion effect, then a clean full reset of the simulation loop.
Object-oriented animation basics: 10–15 fish with wandering behaviour, smooth turning at the tank edges, facing the direction of travel, and delta-time updates. It …
A short but strict game spec: three selectable planes with realistic, displayed performance limits, dogfights against a dynamic number of enemies, visible tracers, …
Design taste and restraint: a luxury single-file site in vanilla HTML, CSS, JS and Three.js where every visual is generated in code, with four required sections and …
Spatial reasoning without seeing the output: the model must place a recognisable pelican on a bicycle with plausible contact points (seat, pedals, handlebars) in raw …
The short original of the heptagon family: collisions with moving walls (the wall's velocity at the contact point must enter the bounce), plus gravity and friction.
Reading versus pattern-matching: the five people on the main track are already dead.
Whether the model recomputes the probabilities when the classic setup is inverted.
Reading the first sentence: the cat is dead when it goes into the box.
Resisting a memorised multi-trip puzzle solution when one trip is enough.
Checking feasibility before searching: every measurable amount is a multiple of gcd(6, 12) = 6.
Avoiding list-writing autopilot on a trivial task.
Units over the memorised 'they weigh the same'.
Adapting to the tool actually given: a digital scale shows weights; it isn't a two-pan balance.
Spotting that the target equals one rope's burn time.
Admitting when the standard tricks can't reach the target.
Reading the negation instead of solving the famous version.
Rate reasoning on a variant of the cognitive reflection test.
Linear versus exponential: the classic lily-pad answer (day 39) assumes doubling.
Detecting a contradiction instead of forcing an assignment.
Common-sense physics hidden in an arithmetic word problem.
Tracking time and gravity in a distracting scene.
Physical intuition: tape on the top surface lifts only the top slice of bread.
Ignoring irrelevant vectors (river, wind, car speed): the glove falls onto the bridge.
Holding a character-level constraint over a long answer; 'c' is everywhere in research vocabulary.
Several independent format constraints at once, plus a forbidden word.
Conflicting pressure between length and sentence count, plus a title format.
Counting lines exactly while keeping punctuation rules.
Choosing the right two of three tools, mapping 'perimeter of a circle' to circumference, passing radius 5 and diameter 10 correctly, and emitting valid JSON only.
Tool-use restraint: the only available function computes BMI, but the user wants a triangle's area.
Public prompts can end up in training data, so a model may have seen them. Use these as a quick, comparable first look, and change details (numbers, names, colours) to check that a model reasons rather than remembers.
New test prompts in your inbox
An occasional short email when strong new test prompts are added, each linked to its source. No spam. Every email has a one-click unsubscribe link.