Source: raw/I_was_a_Claude_hater._Fable_5.2_changed_my_mind..md

Creator: Claire Vo, How I AI (solo “just-in-time mini episode”) · URL: https://www.youtube.com/watch?v=1gYQc4CL_oo · Duration: not recorded · Platform: YouTube (auto-captions, fetched 2026-11-11). No sponsor segment.

Claire Vo had called herself “kind of a Claude hater” since the original Fable and Opus 5, which she found capable but “so annoying to talk to”, and had moved most of her coding to Codex. After a few days of pre-release access to Fable 5.2 she says she is back on Claude Code, mostly with Opus 5.5, and calls Fable 5.2 a “fast, capable, not annoying, creative powerhouse”. This article covers what she built and where it fell short; model facts she relays from the launch (availability, guardrails, benchmarks, pricing) belong in the Fable 5.2 model article. See Claude Fable 5.2 + Mythos 5.2 for the model facts.

Key Takeaways

  • Her pros: fast per turn and fast overall, autonomous but “willing to alert me when it needed something, but no more than that”, strong at creative work, video and animation, 3D and games, architecture and long-running engineering, and “not annoying”. She hit no guardrails in her testing, which she ties to what she tested.
  • Her cons: bad at front end and web design “left to its own devices” (“incredibly concerning taste”), though “really tasteful when you give it a good brand or style guide”. It has no image model, so she kept calling OpenAI’s image gen 2.5 for images.
  • Her split: use Opus 5.5 for daily work and for design; bring Fable 5.2 in for scope, complexity and creativity. She plans to pair Opus 5.5 on the front end with Fable on the back end.
  • Video stopped looking like slides. Her Hyperframes + Opus 5.5 videos “still kind of felt like slides”. Given a tweet about “software factories” plus her company’s cxo.dev brand guidelines and “make this tweet a video”, Fable produced a fully animated story (a team builds a software factory, turns it on, and finds “what they really built is a worse version of Fable Code”, as captioned) with no storyboard or creative direction beyond the brand book. She went back and forth on it, and calls the result “not 100% perfect”.
  • Unprompted sprite animation. The brand’s Space Invader-style bots existed only as static images; earlier models made them “wiggle their way across the screen”. Fable built sprites that walk, cheer and jump. She has added “animate the invader bots” as a test in her How I AI eval bench, where it took 16 minutes.
  • Geometry: flat-pack furniture with no instructions. From low-resolution photos of 25 wooden parts with Chinese-language labels and a bag of screws, it produced a diagram and step-by-step assembly in the right order. She says every other model she tried the week before failed, and the assembler followed Fable’s instructions.
  • A funny game. “Make a game about fighting bugs using our Space Invaders” produced a shooter where you stop bugs before production, with trade-offs (more agents catch more bugs but let more through), a trailer joke about finance downgrading the team to “the cheap model”, and the monorepo as final boss. She says it made her giggle.
  • Real engineering on a messy codebase. Pointed at ChatPRD’s product-intelligence pipeline (ingest, organize, summarize, auto-wiki; six months of iteration) with “make the pipeline more efficient, easier to maintain”, it produced a plan that matched her own roadmap, worked autonomously, and stopped “at the right moments” until it needed a PR. She was about halfway through the migration at recording.
  • Bad web design, shown. A “slow down” browser homepage (“so ugly, so much slop”) and a ghostly storybook (“okay, but pretty ugly”).
  • The highlight: a one-shot 3D typewriter. Inspired by a 3D record player from “Guillermo at Vercel” (presumably Guillermo Rauch ^[inferred]; he published the source), she asked Fable for a similar but different object. One shot plus later feature requests gave her a 3D typewriter you type on by click or keyboard, with whiteout and overtype, rain and music, 15- and 30-minute sand-glass writing timers, writing prompts, the record player integrated as a radio, scenes (soft evening that follows the time of day, a train, an oil lamp, handwritten pages with a pencil, a Y2K computer with green or amber screen), and dot-matrix printing into saved stacks. She plans to deploy it.

Try It

  1. For a video, give the model your brand guide and one source artifact (a tweet, meeting notes) and ask it to make the video, before adding any storyboard. Compare against your current Hyperframes or slide-style output.
  2. Test spatial reasoning on a real object. Photograph the parts of something with missing instructions and ask for an ordered assembly diagram. Check the order before you start building.
  3. Give it a style guide for anything visual on the web. Her experience: left alone its taste is poor; with a brand guide it follows well.
  4. Route by task: Fable 5.2 for long refactors, architecture and creative builds; Opus 5.5 for daily work and front end; a separate image model for images.
  5. Build your own small benchmark from tasks that failed before, the way she added the invader animation and the furniture instructions. A task every model failed last week is a better test than a public benchmark.

Open Questions

  • The captions render the desktop app she had switched to as “Claude X”; from context (browser and computer use “light years ahead of Claude Code”, matching her earlier Codex coverage) it is read here as Codex.
  • The full How I AI bench comparison against other models was promised for “a week or two” later and is not in this episode.
  • The typewriter deploy URL is not in the transcript.
  • Model facts she relays (the Mythos 5.2 companion model, the black-hole physics result, benchmark names, availability channels, a fast mode, subscriber resets, cyber/bio guardrails and anti-distillation safeguards) are sent to the Fable 5.2 model article for verification. A distribution surface captioned “Cloud Kube” could not be identified.