Fable 5.1 and 5.2, one design review
· by Jonas
A design kit for my sites and tools has a list of settled decisions: 26 of them, each with the words that settled it. The kit covers a dashboard for ML training runs, a personal site, a file-converter page and public project trackers. Fable 5.1 and Fable 5.2 each reviewed it from the same input, to see how the newer model reads the same work.
The setup#
- Same folder: the list of settled decisions as text, and the same 21 stills of the pages and the kit's own documentation. Nothing else.
- Same brief, word for word: "Write a walkthrough review of this design kit for a reader who hasn't seen it. Go page by page: for each still, write a short paragraph on what works and what doesn't, with specifics. Then give your verdict on each settled decision (keep, revisit, or cut) with the argument for it. Finish with what's missing. Write it as an essay, in plain prose." One more line named the files and their order.
- Same harness: Claude Code run headless, with no instructions file, skills, plugins or agents, and read access only. One run each.
| model | words | turns | time |
|---|---|---|---|
| Fable 5.2 | 2,218 | 32 | about 4 min |
| Fable 5.1 | 3,195 | 31 | about 6 min |
How each wrote#
5.2 wrote in headed sections, with lists for problems and a verdict on each of the 26 decisions under its own number. It backed claims with measurements: digits that "break apart" as "0.69 17", a takeover that "costs about 0.8 seconds per click". It also said what it couldn't judge. Motion was judged "from two contact sheets of small frames", and four parts got no verdict because they "appear in no still".
5.1 wrote continuous prose, a paragraph per still, with its verdicts grouped into long paragraphs. It is about half again as long, and it argues every verdict in full sentences.
Where they agreed#
On most of the big calls:
- Readings: the dashboard sets every number in a seven-segment face reserved for live values, on a run that is finished. 5.2: "The run is finished and the page says "Nothing is live right now", yet every reading is in the seven-segment face that is reserved for live values."
- Dark mode: the dark side of each page flips to cream. 5.1 put it as "a training dashboard is the kind of page someone watches at night".
- Red: both counted far more uses of the signal colour than the kit's rule allows, and both would take it off link arrows and counts.
- Unused marks: both would move marks no page uses out of the kit.
- The best part: both named the dashboard's run list the most resolved component.
- What's missing: both put problem states first, a failed run or a rejected file. 5.1: "A dashboard for training runs will need one on its first bad night".
Where they split#
- The full-screen takeover on load. 5.2 would cut it: "Choosing a run or dropping a file reveals something. The landing commit reveals nothing new." 5.1 would keep it on load and shrink it for clicks in a list: "Landing earns a takeover."
- The loss line crossing the page's diagonal. 5.2 called it "the best single detail in the kit". 5.1 read the same detail as a sign that "data, unlike text, has no rule about crossing".
- The reveal on load. 5.2: "It is the reveal, and it should be the whole landing." 5.1 saw a pink wash in the first frames: "Either the pink is the intended unscanned tint and the kit should say so, or it is a bug."
What each caught alone#
- 5.2: the kit's own page contradicting itself on one mark, missing-glyph boxes on the converter page, and a three-digit run number as an untested case.
- 5.1: one shape carrying two meanings on the trackers, class names a new project couldn't read, and a minified theme file.
What it shows, and what it doesn't#
- 5.2 is the easier review to act on. It is shorter, numbered, measured, and careful about what it saw.
- 5.1 notices more and argues more, and it is surer of itself where it saw less. It ruled on a part that appears in no still, from the decision list alone.
- They agree on more than they split. Where they split, both cases are argued well enough to be worth reading.
It is one brief, one input and one run per model. Read it as two worked examples, not a ranking.
The transcripts#
Every message, tool call and still, redacted, is on baldanders: Fable 5.2's review and Fable 5.1's review. One still is cut from both, a sheet of the kit's marks on their own.