Grok writes
Grok writes
Verdict: Grok 4.6 is the default writer for this vault. Wiki pages, journal, decisions, and reports go to Grok. Banks and local execution stay on Grok. The 13–15 August ranking that put Fable on every page this desk has to like is retired for routing. It stays on disk as the test that was true then.
What was being decided: who writes prose this desk has to like, against leaving that seat on Claude Fable 5 (Opus when Fable is out).
What flipped
On 13 August, three wiki pages were regenerated from identical banks. Fable won all three. The owner’s line was “not even close.” On 14 August, 52 Grok intros failed the pillow test and Fable took the doors. On 15 August the stack was signed: Fable writes, Grok 4.6 banks and executes. The written flip condition was a Grok slate ranked above Fable on three pages, or Fable’s trust premium dying on the pages that stayed in Cowork.
On 1 September the owner said Fable’s writing has been poor lately, that good quality now takes a lot of tokens, that Grok has stayed consistent, and that Grok should do the writing. That is the trust-premium death the 15 August note named. It is not a new three-page bakeoff. It is the product in use.
The public record in State of AI, end of August 2026 already had Claude’s default English as the loud failure of the month. Fable is the same lab family. The dialect, the rewrite tax, and the token burn are one problem.
Steelman for keeping Fable
Fable still wins blind code-quality reviews. Arena writing boards still put it at the top. This vault’s own A/B/C on 13 August is the only scored writing test this desk has run, and Grok lost it. Grok’s known miss here is compressing an operating claim into a quote card, then shipping. Holdings.py, the generator, and the owner’s eye on the door exist because of that miss. Handing Grok the writer seat without those checks would replay mid-August.
Keep Fable if the next three pages Grok ships need a rewrite tax as heavy as Fable’s current one, or if a new A/B/C on live pages ranks Fable first again. Until then the 13 August win is a dated score, not a live product.
Cost of this seat
Grok still has to pass the same filters: generator, Writing Standards as a gate after the draft, holdings.py, regen-check on wiki regen, owner eye on openings. The 52 failed intros are not forgiven. They are why the filters stay.
What this ruling does not cover: tsumugu companion fiction and dialogue. That lane has its own authoring gate and was not scored today. Cowork can still research. It is not the default writer.
Live routing
| Job | Seat |
|---|---|
| Wiki prose, journal, decisions, reports | Grok 4.6 |
| Research banks, vault scripts, local execution | Grok Build calling 4.6 |
| tsumugu-core / tan / frontends | Cursor |
| Standing duty, packet only, public only | Grok Bot |
| tsumugu companion fiction and dialogue | Unchanged until scored |
Pairing that earns rent: Grok banks, Grok writes, checker grades, owner eyes the door. Do not send the draft through Fable “for polish” as a default. That is the token tax this ruling exists to stop.
Related
- What works: Grok 4.6 and Grok Bot — the ranking this note retires for the writer seat
- A/B/C writing test: Fable takes all three — the test that was true on 13 August
- State of AI, end of August 2026 — the public month that made the token tax visible
- Current Agentic LLM Stack — the live roster
Decided / outstanding / next
Decided: Grok writes vault prose. Fable is off the default writer seat.
Outstanding: tsumugu companion fiction. Whether a later Fable checkpoint earns the seat back.
Next move: write here first. Do not open a Cowork window to “fix the English.”