Cursor Ultra vs Grok Build vs Grok Bot

You have three unused-or-underused seats on one Heavy bill. Cursor Ultra is at 0%. Grok Build is at 21%. Grok Bot comes with Ultra and you are not using it. They do not share a token bucket. The wiki and the codebases already have a split: Grok Build and Claude Code do the vault and batch authoring; nothing is using the IDE seat.

Put Cursor on the TypeScript/Swift/HTML products. Keep Grok Build on the vault, research, and batch pipelines. Use Grok Bot for work that should keep going when you are not at the TUI.

What you actually run

Project	What it is	Shape	Who should own it
llm-knowledge-base	Karpathy wiki + Astro site + regen + journal	355 wiki pages, 887 raw notes, 124 journal files, research banks	Grok Build (and Claude Code). Not Cursor as the writer.
tsumugu-core	Offline reader, voice, blog, workers	~2000 app files, TS/Vite, lots of audio	Cursor Ultra for code. Build for content/docs.
tsumugu-ed	Dictionary: 9,663 entries, example-sentence factory	10k+ JSON, Python/TS work orders, audio still thin	Build for authoring loops. Cursor for the site/frontend PRDs.
tsumugu-wiki	Generated language wiki	Huge compiled tree	Same as the vault: Build/Claude, not Cursor-as-author.
tan	iOS audio notebook. PRD unsigned except yuedu. Swift v0 skeleton	app/Tan.xcodeproj, library pipeline not started	Cursor for the Swift app. Build for Mac-side TTS/library scripts.
wnac	Finance dashboard, P1 live	Python → dashboard.html	Cursor if you restyle or extend the dashboard. Build for analyze.py / insight prompts.
cos	Personal OS, HTML dashboard, TUI retired	Deprioritized	Leave it. Bot only if you revive daily refresh.
wnab	Actual Budget fork, retired	Archive	Do not open in Cursor “to use Ultra.”
kind-travel-mockups	Small HTML	Side	Cursor if you touch it.

The wiki already follows Karpathy: raw/ immutable, wiki/ compiled, schema in AGENTS.md. Cursor does not make that better. It makes seeing and changing application code better.

What each product is

Grok Build (this TUI, 21%). Agent in the terminal. Files, git, subagents, workflows, X search, the study lane, open to Obsidian. Already how tsumugu batch work and the knowledge base get done. Best when the job is “read many markdown files, write markdown or JSON, run a script.”

Cursor Ultra (0%). An IDE with Composer and Grok 4.6 (and other models) on a separate meter. Diffs, multi-file TS/Swift, jumping a 2000-file app, UI you can click. Same model family as Build, different harness and different policy (Cursor’s Grok is tighter on cyber-ish tasks; Build is the looser coding agent).

Grok Bot. Desktop/mobile agent that keeps working when the TUI is closed. Knowledge-work coworker, not a wiki maintainer. Included with Ultra. Do not give it write access to wiki/ without the schema and your eye.

Routing

Job	Where
Wiki ingest, regen, journal, research banks, lint, writing-standards work	Grok Build
Tomorrow’s skills / intro study lane	Grok Build (two windows)
tsumugu-core reader, Vite, tests, worker, UI	Cursor Ultra
tan SwiftUI app, Xcode project, iOS build errors	Cursor Ultra
tsumugu-ed example-sentence authoring (7,001 words still ~0%)	Grok Build (batch, already the production line)
tsumugu-ed / tsumugu-core frontend bugs, layout, CSS	Cursor Ultra
tan library generation, TTS, manifest.json on the Mac	Grok Build
wnac dashboard.html restyle, JS tabs	Cursor Ultra
wnac categorization / brain.py prompt work	Grok Build
Overnight “keep researching X / draft a brief I’ll file tomorrow”	Grok Bot
Daily tan yuedu source hunt (X, news) while you are not at the desk	Grok Bot → files in a drop folder; Build compiles into the vault/library
Anything that must match Writing Standards or report intros	Grok Build, after the how-i-write skill exists
Multi-model compare inside one editor (Composer vs Grok 4.6)	Cursor Ultra
Click around a giant TS repo and ask “why is this red”	Cursor Ultra

Hard no for Cursor: do not make it the wiki’s author. You already decided the model does not own the prose. Cursor’s 0% should go to tsumugu-core and tan, where an IDE is the point.

Hard no for Bot: do not let it ingest into wiki/ unattended. Drop into raw/inbox/ or a tan staging folder. Compile on Build, with you looking.

First uses of the 0% (so Ultra is not wasted)

Do these in Cursor, Grok 4.6, Ultra meter:

1. Open /Users/n1/Projects/tsumugu-core. Run the app’s typecheck/tests. Pick one real UI bug or the next reader feature. Stay in the editor for a week of that repo.
2. Open /Users/n1/Projects/tan/app. Get Tan.xcodeproj building. Cursor for Swift; do not move the unsigned PRD into “the model decided.”
3. If the dictionary site still has the frontend repair PRD, do that in Cursor against whatever repo serves the public site — not by regenerating 10k JSON.

Leave Grok Build on: vault, tsumugu-ed authoring, tan content pipeline, this study lane.

Leave Grok Bot for one experiment: a drop-folder brief (X + news → a markdown file in raw/inbox/ or tan library/ staging). You file it. Build does not have to be open.

How this fits the study lane

Stage 2 skills stay on Build. /how-i-write and report intros are for this TUI and Claude Code, not for Cursor Composer.

If Cursor needs a voice, that is a Cursor rule (.cursor/rules) copied from the same journal notes — a second copy of the law, not a second skill ecosystem. Rule that in Stage 2 if you want one file in both harnesses.

What not to do

• Do not “use Ultra” by opening random markdown so Cursor is the default .md app. That is not leverage; that is Launch Services.
• Do not run the 7,001-word example-sentence factory inside Cursor. It is a batch authoring line. Build already does that.
• Do not buy a fourth coding agent. You have Build, Ultra, and Claude Code. Bot is the away-from-desk slot.