Claude Fable 5 claims 8,000-line Pokémon game from one prompt

• Anthropic pitched Fable 5 as the first public Mythos-class Claude in its launch post, and the first demo wave immediately centered on browser-native games, sims, and creative coding projects that creators said they built from short prompts or goal-based specs, from bilawalsidhu's city block simulator to venturetwins' AI Monopoly.
• The headline example was a claimed 8,000-line Pokémon game from one prompt, which minchoi's Pokémon example amplified, while om_patel5's FireRed post went further and claimed Fable 5 beat Pokémon FireRed start to finish from raw screenshots alone.
• The demos were not all pure one-shots: petergyang's F-Zero test needed follow-up prompts for speed feel, petergyang's sprite-sheet reply said retro asset alignment still trips models up, and venturetwins' follow-up described AI Monopoly as under 20 minutes of back and forth rather than a single untouched pass.
• The workflow shift showed up in the official docs too: Prompting Claude Fable 5 says the model is built for long-running end-to-end work and parallel subagents, while ClaudeDevs' setup thread said prompting gets simpler, thinking is always on, and older prescriptive skills can undersell it.

You can read Anthropic's launch post, skim the official Fable prompting guide, check Every's early vibe check, and even open Glenn Has A Beard's public platformer artifact. The weirdest claim was still om_patel5's FireRed thread, but the more useful reveals came from the smaller posts that showed where short prompts held up, where image references helped, and where creators still had to steer.

The first wave landed like Christmas for vibe-coding people because the examples were not todo apps. They were playable worlds with systems, camera logic, and enough polish to survive a public demo.

According to bilawalsidhu's post, Fable one-shotted a city block simulator with multi-agent traffic, live detection boxes, tracks, and a day-to-night cycle. In a reply, bilawalsidhu's prompt note said the setup was a "/goal" plus a "chonky well defined prompt," which lines up with Anthropic's own guidance around explicit outcomes and longer runs.

venturetwins' AI Monopoly and om_patel5's COD Zombies clone showed the same pattern in two different directions: one sentence or one themed prompt, then a browser game with rules, multiplayer codes, or FPS gunplay already wired up. Venturetwins said the Monopoly build covered game rules, money, turns, share codes, and even data-center upgrades on owned properties.

Some of the strongest creator demos were not about raw one-shotting. They were about getting a model to hold onto a visual identity.

Glenn Has A Beard used reference images for characters in a simple fighting game, then fed Fable a batch of recent artwork so it could apply that look to a platformer in the first test and in the follow-up. The artifact linked from that thread is public as CLIMB X, and Glenn added that it also ran on mobile in the mobile reply.

That is a more concrete creative jump than generic "make me a game" demos. The model is not just scaffolding mechanics, it is also getting pulled toward a house style from image references and then carrying that style into something playable.

Short prompts were real, but the more repeatable wins came from people who packed a lot of constraints into a single ask.

Peter Yang's F-Zero test is the clearest example because his posted prompt reads like a compact design brief, not a magic phrase. It specified pseudo-3D track rendering, AI opponents, a boost meter that drains health, visible speed cues, a HUD, checkpoint logic, and a neon cyberpunk style.

The early Fable playbook that emerged across bilawalsidhu's /goal note, ClaudeDevs' setup thread, and the official prompting guide looked like this:

• Set a clear goal or outcome.
• Give the model a task that would normally take hours or days.
• Pack concrete mechanics and constraints into the first prompt.
• Let it run longer instead of steering every step.
• Come back for targeted fixes, not constant co-piloting.

That is close to how Anthropic framed it in docs and in Every's early write-up, which described Fable as strongest when treated like an async whole-job agent rather than a chatty collaborator.

The demos were impressive, but the most useful posts were the ones that admitted where the first pass still missed.

Peter Yang said Fable did not quite one-shot the F-Zero build and needed extra prompts like "give it more of a sense of speed" in the main test. He also said sprite sheets still tend to break on alignment and sizing in a follow-up reply, which is a very specific failure mode anyone making retro-styled games will recognize.

Venturetwins added another useful calibration point. In a reply about AI Monopoly, they said the result had no errors, but still involved back and forth for tweaks like slower animation and a monopoly tracker, with the whole session taking less than 20 minutes.

Even the backlash posts were mildly informative. AIandDesign's skeptical post argued the X hype was overselling things, which felt fair given how many examples mixed true one-shots with short iteration loops.

The Pokémon material dominated the launch-day feed because it bundled three different claims into one memeable story: long horizon planning, pure vision, and an end-to-end completion people could immediately understand.

Om Patel claimed Fable 5 cleared Pokémon FireRed in 50 minutes of in-game time with no minimap, coordinates, memory hacks, or external tools, just screenshots of the game state in the original thread. Min Choi helped spread that claim in a shorter repost, and then folded a separate Pokémon game build into his 10-example roundup.

The important split is that beating FireRed and generating an 8,000-line Pokémon game are not the same demo. One is a long-horizon vision-agent claim, the other is a code generation claim. Both fed the same narrative, which is why they blurred together so quickly on launch day.

The last useful reveal was not in the games. It was in the product behavior around them.

Anthropic's model docs say Fable 5 ships with adaptive thinking always on, a 1 million token context window, and safety classifiers that can return a refusal stop reason or reroute certain cyber, bio, and related requests to Opus 4.8. In the public setup thread, ClaudeDevs said those reroutes hit under 5% of sessions on average.

That helps explain why so many early creators described the model as slower, more autonomous, and better suited to heavyweight runs than interactive tinkering. In the same thread, ClaudeDevs said older prompts and skills can be too prescriptive for Fable, while the official prompting guide says teams are getting the best results by giving it harder end-to-end work, using parallel subagents, and lowering effort only when the model starts over-deliberating.