How work gets decided, who does it, and which wallet pays. This is the half of the plan that does not decay. What is actually in flight lives on the Board, and what is waiting on you lives under Decisions.
01The Fact That Changed Everything
2026-07-12
Two of four proposals bet the week on building the Salapaw funnel pack. It already existed, finished, on disk. Nobody had looked.
Check the Disk Before Proposing Work
Verify delivery state first. This has now cost a full planning round twice. It is the reason every plan in this studio starts with a read, not an idea.
02The Roster
Not a single assistant — a roster, each with a designated role and a wallet it burns.
| Model | Role | Lane |
|---|---|---|
| ☼ Sonnet The Step-Ahead |
The hands, and the default daily driver. The executor subagent on every bounded build, plus the independent pr-reviewer pass. | Workshop |
| 🔆 Opus The Architect |
Plans every handoff, reviews at milestone boundaries, sole merge authority. Not the everyday model — running Opus all week is what hits the rate limit. | Workshop |
| 🦋 Fable The Far-Seer |
Long autonomous runs only. Opened deliberately, never by default. | Boardroom |
| 🌞 Sol · 🌏 Terra | On ChatGPT Go, which has no Codex — conversation partners, not hands. | ChatGPT Go |
Retired: ⚙️ Composer 2.5 (2026-08-26 — its role passed to Sonnet as the in-session executor) and 🤖 Grok 4.5 (2026-08-21, with Cursor). Grok invented the lane discipline on this page. The rules outlived the model.
03Token Lanes — Three Wallets
Cursor was cancelled 2026-08-21 · review date 2026-09-19
The old four-lane map was built on Cursor’s two wallets. Both are gone, and so is the usage panel that used to settle which lane a model belonged to. On the way out it showed the included lane 13.9% used across a whole month while the metered one sat at 100%, emptied largely by a single model. That asymmetry is why cancelling was the cheap move.
🔧 Workshop Claude Pro — the main lane
Everything that builds: planning, the whole handoff loop, reviews, audits, specs, migrations. Sessions reset on a rolling window — and it is now the only lane that recovers on its own.
Manage it: Sonnet by default, Opus for plans and merge verdicts. Batch a session’s worth of handoffs rather than trickling them.
💎 Boardroom Claude usage credits, prepaid — Fable only
Roughly $10/M in and $50/M out — twice Opus, against an eval record where Opus wins 7 of 8 head-to-head. Every sitting is a deliberate purchase.
The one question that decides it: is this one long autonomous run, or a good document something else will execute? Fable earns its price on endurance across days, not on being smarter per question. Using it to write a plan another model executes throws away the part you paid for and leaves an expensive planning document.
🛋️ Salon Claude.ai
Life planning, narrative, synthesis — where the conversation itself is the product. Also where found this captures start. Walk in with a brief, walk out with a deliverable.
🏭 Factory — retired
Cursor first-party. Held Auto, Composer 2.5 and Cursor Grok 4.5. Died with the subscription; execution had already moved to the Workshop. Cursor Cloud agents went with it — there is no external dispatch lane any more.
04Dispatch — GitHub Is the Switchboard
All cross-model work moves as Issues and PRs on the repo where the code lives, aggregated on the FoundThis project board. Opus reviews everything before it reaches Donsan.
1 · Plan
🔆 Opus writes a bounded issue: goal, build, acceptance criteria, constraints, and what not to touch. Self-contained, because its next reader will be cold.
2 · Execute
Opus briefs the in-session ☼ Sonnet executor with what the issue could not carry — the reasoning, the approaches already ruled out, the disk state. It branches, implements, verifies by running, opens a PR. It never merges.
3 · Independent Read
A fresh Sonnet pr-reviewer posts one advisory comment. Never the agent that built it: since planner and executor now share a session, this is the only reader in the loop that has not seen the plan.
4 · Verdict
🔆 Opus checks every acceptance criterion, merges or requests changes.
Every Issue Needs Five Things
Files to touch with full paths · the signature being added, with types · the schema or query written out, not left to guess · acceptance criteria that are checkable, not “works correctly” · and what not to touch, to bound scope creep. Missing any one of them means the planner is not done. Do not hand off yet.
05House Rules
- Tokens spent is not success. Published work, removed uncertainty, closed loops.
- One vertical per day. Stop when shipped or honestly blocked. Do not fake a session.
- If the deep task overruns, everything below it drops. One finished loop beats five open ones.
- Check the disk before proposing work. See section 01 — this rule was paid for.
- Opus owns code decisions. Donsan reviews product direction, UI and copy, external surfaces, and anything genuinely blocked. He does not review code.
06Refuse on Sight
Work that feels productive and is how a surplus week evaporates with nothing published.
- Opening work because tokens are left over
- FoundThis platform features before the Vault is validated
- An AsianWall marketplace
- A peace4life paywall before retention exists
- Parallel builds across cold verticals
07Where Things Are Written Down
| What | Where |
|---|---|
| Canonical roster + lanes | .claude/models/workflow.md |
| Active plan, workstreams, sequencing | MASTER-PLAN.md |
| Handoff mechanics | .claude/skills/handoff/SKILL.md |
| Why a decision was made | the Vault, under its lens |
| What to build because of it | the repo doc it governs |
Vault or Repo
The vault holds why. The repo holds what. A builder working in a repo must never have to open the vault to build correctly — so when an insight implies a decision, do both, and never leave a build-affecting decision in the vault only.