On this page
- The outcome we are organizing around
- Current evidence and corrections to the old plan
- Five seats, five owners
- Building program: all 13 families
- Atlas-to-scene coverage
- A package Quik can actually build
- Readiness states
- Parallel work and model use
- Branches, claims and one integrator
- Checkpoints and the rolling buffer
- First launch
Crown Key — five-seat production plan R02
Prepared 2026-10-17 UTC / 2026-10-16 America/New_York. Baseline: 215GrandAve/lost-souls main 4ce28d89b6fb1922f0b8427937b986961504307d.
This is the new creator-directed production program: each enclosed Crown Key building receives a usable entrance/lobby and one or two purposeful rooms, supported by a coherent exterior package. Existing Crown Landmark work is preserved. Open pavilions, canopies and infrastructure receive appropriate usable spaces and interfaces. This replaces the R01 recommendation to defer most interiors; it does not retroactively mark new work completed or accepted.
Active production run: packages, images, blueprints and materials records the actual in-session candidate work and review findings. Its local subagents are separate from the external-seat launch briefs below; their work does not imply that another account's seat or editor is running.
The outcome we are organizing around
Quik has sent his main Claude the first-session brief. Keep that assignment: orientation fixture → connected Unreal sample → isolated Crown Landmark review/modeling. Shared-world placement and district construction remain held under that brief. Our first delivery is Split Complex ready for him to build, with Terrace Tower developed concurrently as the fallback and second ready package. Frame Tall and Crescent Hotel follow. The operating target is always two build-ready buildings ahead of implementation; the verified ready buffer currently contains zero next-building packages. Concepts, scope sheets and generated images do not count as ready inventory.
The five seats work on different deliverables. OA1 integrates and delegates independent release review to read-only reviewers; OA2 and OA3 produce complete building packages; Claude1 makes the deterministic pipeline useful for those packages; Claude2 continues Quik's actual engine work. Quik’s Claude is already briefed; its actual execution is unverified here. Its incoming-package feedback is useful when available, but is not a required gate for the next building. The other external-seat briefs are ready to launch, not claimed running sessions.
Current evidence and corrections to the old plan
- Crown Landmark EXT R06 and INT R02 remain issued reference sources. Quik's new implementation work can proceed in an isolated working branch/map. Record what he actually imports or models; do not describe it as engine-accepted before the evidence exists.
- Crown Landmark PR #45, including C02 and the 81-room fit work, is merged at
a7220db. Read the merged files and current issue list. Preserve C01 history; do not use an old open-PR description as current status. - Tooling PR #51 is merged at
7a81eb8. Extend the actual merged implementation rather than creating a second competing pipeline. Its test-only evidence does not establish a successful Crown Key production import. - The Claude CB00 Split study has two bridges; the latest H1 concept direction has one enclosed bridge. OA2 must reconcile this discrepancy into one selected topology record before Claude1 adapts a production candidate.
- World placement, coordinate frame, shoreline and connection decisions still require evidence. This is not a blanket pause on isolated asset modeling, design development or staging.
- The two recent Terrace/Frame images are concept candidates. Preserve their provenance and evaluate them against the controlling catalog; they are not measured views from a shared model.
Five seats, five owners
| Seat | First ownership | First concrete delivery | Follow-on |
|---|---|---|---|
| OA1 — this Codex seat | Integration, atlas coverage, independent release review, shared plan and .world publishing | All 31 assets have a current scope; atlas scene requirements have source/status/owner; read-only reviewers check geometry/routes/packages | Record review evidence; merge reviewed packages; publish asset pages; keep the two-building queue honest |
| OA2 — OpenAI asset lead A | Split Complex PI-BLD-0002-V01 | One topology decision record; exterior/entrance/lobby/gallery/receiving-space package; dimensions and route evidence tied to its editable candidate | Crescent Hotel, Compact Terrace Block, then Arcade/Pavilion/Canopy families |
| OA3 — OpenAI asset lead B | Terrace Tower PI-BLD-0003-V01 | Coherent exterior and residential entrance/lobby/studio/terrace-lounge package, with editable candidate and schedules | Frame Tall, then Frame variants; Twist/Garden/Office/Residence/Straight families by atlas priority |
| CLAUDE1 — production tooling | Reuse merged PR #51 pipeline for the first two assets | Reproducible export/drawing/manifest package from selected inputs; an actionable Quik handoff rather than another isolated demo | Fix pipeline blockers once, then apply to subsequent owned assets |
| CLAUDE2 — Quik’s existing main developer | Existing fixture/sample/Crown execution | Actual import/build feedback from Quik's current work | Incoming-package and engine feedback as available; no required next-building release-review duty |
Legend owns creative choices and disputed canon/scene assignments. Quik and his main Claude own local implementation and actual engine observations; only one agent controls the running editor at a time. OA1 owns the integrated ledger, independent review records and release decision. OA1 delegates geometry, route and package checks to reviewers who did not author the candidate and have read-only access to it; use Astra for a bounded difficult review. Claude2 does not independently approve its own Crown work and is not the review bottleneck for Split or Terrace. A technical reviewer can reject an inconsistent package; they cannot silently change its creative program.
Building program: all 13 families
Room uses below are working production choices, not newly established story canon. OA1's atlas audit may add or change a room for a supported scene. Every variant retains its own ID, entrance/interface check and coverage record even when it reuses a room kit.
| Family / variants | Lobby or arrival | One or two useful rooms | Queue/owner |
|---|---|---|---|
0001 Crown Landmark V01 | Preserve issued entrance, reception and atrium | Preserve existing council/diplomacy and other issued interiors; reconcile scene coverage and fit issues | Quik now; CLAUDE2 support |
0002 Split Complex V01 | Primary tower lobby and core connection | Enclosed gallery bridge; receiving lounge/overlook | First — OA2 |
0003 Terrace Tower V01 | Residential lobby | Furnished studio; residents' terrace lounge | Parallel fallback — OA3 |
0004 Twist Tall V01 / Compact V02 | Sculptural lobby appropriate to each variant | Private office; meeting lounge | OA3 queue |
0005 Garden Canyon V01 / Terraced Garden V02 | Garden-facing lobby | Furnished studio; shared garden lounge | OA3 queue |
0006 Frame Tall V01 / Mid V02 / Compact V03 | Exoframe/core entrance lobby | Meeting room; private office | Frame Tall next — OA3 |
0007 Courtyard Offices V01 / Glass Office Block V02 | Reception and waiting | Workspace sample; conference room | OA3 queue |
0008 Stepped Residences V01 / Compact Residences V02 | Residential lobby | Furnished studio; residents' lounge | OA3 queue |
0009 Crescent Hotel V01 / Compact Terrace Block V02 | Hotel reception; smaller reception for V02 | Guest room; hospitality/terrace lounge, subject to V02 use confirmation | Crescent next — OA2 |
0010 Long V01 / Curved V02 / Corner V03 Arcade | Vestibule or covered arrival | One shop/cafe; its service/back room | OA2 queue |
0011 Waterfront V01 / Inner V02 Pavilion | Sheltered arrival and usable public area | Lounge/kiosk; small service/store where an enclosed volume is intended | OA2 queue |
0012 West V01 / East V02 Arrival Canopy | Open arrival concourse | Information/control booth if needed; do not add an artificial second room | OA2 queue |
0013 Straight Tall V01 / Straight Slim V02 | Clearly identified shared lobby | Private office/workspace; meeting lounge, pending atlas/use check | OA3 queue |
The six PI-STR assets receive usable access and interface packages: Bridge Span, Bridge Landing, Seawall/Promenade, Finger Dock, Dock Connection and Arcade Bay Module. Document movement, edges, level transitions, collision intent, lighting/materials and connections. An arcade module supports the associated room bay; a dock does not acquire a fictional lobby.
OA1 may reprioritize the queued families after atlas findings. Keep Split and Terrace in progress until a specific blocker or a more urgent supported scene justifies switching. Do not begin thirteen architectural studies simultaneously.
Atlas-to-scene coverage
For every relevant atlas location/scene, record: atlas ID and exact source, era, confirmed or proposed Crown Key assignment, building/space candidate, required action and circulation, distinctive props, view/time-of-day needs, existing references, missing deliverable, owner and status. Explicitly track unresolved location assignments and scenes outside Crown Key. Do not move another district's scene to Crown Key to fill a room.
Separate an architectural room from its scene dressing. Several scenes may reuse the same room with different props, lighting or access state. Generate missing spaces and necessary scene views after this mapping; do not count unrelated beauty images as scene coverage. Preserve the existing proposed status of Crown Landmark story-room mappings until the controlling sources or Legend resolve them.
A package Quik can actually build
Every building package has a short START_HERE, exact asset ID/revision, controlling references, chosen uses, build sequence, open issues and a file manifest. The package connects the outside to the inside through an explicit route. It contains:
- Exterior coverage: full-building identity, opposite side, entrance and defining detail. Reuse valid existing images; fill missing evidence rather than generating duplicate angles for a numerical quota.
- Interior coverage: lobby and the selected one or two rooms, with consistent doors, windows, vertical circulation, materials and exterior context.
- A single dimension source of truth with units, local axes, origin/datum and proposed/measured/accepted status. Unknown site measurements stay unknown. Plans, sections, elevations, opening/door schedules and room schedules derive from the chosen candidate rather than contradicting it.
- Editable blockout/candidate, documented exports, named objects/material slots, and saved cameras for the spatial views. Image-generation concepts are marked as concepts until checked against this geometry.
- Materials, repeatable components and asset/prop list; closed or nonplayable routes stated explicitly. Reuse a reviewed kit when its footprint/core/opening requirements are compatible.
- Review findings, import instructions, actual tool versions, tested checks and remaining risks. Packaged files and website links must resolve.
Do not author an entire office tower floor by floor. Build the usable slice, its exterior and the structural/circulation interfaces needed to implement it coherently.
Readiness states
| State | Evidence required | What it permits |
|---|---|---|
| Scoped | Program, controlling references, atlas needs, owner and missing work are explicit | Focused design work |
| Design-ready | Selected visual/topology direction; coherent exterior/interior route; room/scene coverage; unresolved choices listed | Candidate geometry and technical coordination |
| Build-ready | Coherent editable candidate, dimension contract, drawings/rooms/openings/materials, reproducible exports, independent review and usable handoff; no unresolved blocker in the intended build slice | Quik can start that asset without inventing its basic design |
| Engine-verified | Actual project evidence for scale/orientation, import/materials, collision/traversal and agreed runtime checks | Claim the specific tested implementation is verified |
Build-ready does not mean engine-verified or finally placed in the district. Mark a packet as partial when only part of it qualifies. An unresolved Split bridge count is a blocker to its complete build-ready designation.
Parallel work and model use
Use Sol 6.1 for the ordinary asset-owner, source-audit, implementation and packaging loops. Use Astra for bounded hard reviews: competing atlas evidence, Split bridge topology, shell/room-fit conflicts or a high-impact release review. Its task ends with a decision matrix or defect list. This is task routing, not a pricing or entitlement claim; use the models/tools available in each account.
Each asset seat should run at most three useful child tasks at once: source/atlas extraction, a separately owned interior/prop draft, and independent consistency review. The lead owns the final coherent package and dimension contract. Keep writers in disjoint paths; reviewers read only. Start small and expand only when independent work exists. Seven available agents in this session is a ceiling including the lead, not a requirement to keep seven busy. External seat sessions and Claude subagents have their own actual limits.
The image tool produces images; a language-model subagent does not become another image-generation entitlement. Queue distinct shots with fixed references and image records. Use editable geometry and saved-camera renders for spatial consistency, then concepts for appearance development where useful.
Branches, claims and one integrator
Start from the pinned baseline, then fetch current main and read AGENTS.md, current decisions and the current protocol pointer before edits. Reconcile any changes since the pin. The new workstreams use separate branches/worktrees: proposed branches are codex/ck-r02-integration, codex/ck-r02-split, codex/ck-r02-terrace, and claude/ck-r02-tooling. Quik's existing Claude preserves its current execution session and branch/worktree; this addendum requires no new branch or restart.
The seat briefs name initial exact path claims. They are proposed assignments until the seat acknowledges its baseline/branch and OA1 records the active claim. A claim contains owner, paths, base SHA, status and deliverable. Shared schemas, generators, root navigation, shared catalogs and built website output belong to one owner at a time. OA1 alone merges, rebuilds shared pages and publishes .world. No .com work is included.
An asset owner does not edit another asset's files. Claude1 initially owns docs/crown-key-production/R02/claude-tooling/ and identifies the exact existing tooling files before changing them. Tooling exports go into the asset owner's agreed handoff path through a reviewed handoff, not competing edits. OA1 owns docs/crown-key-production/R02/reviews/ and records findings from its delegated read-only reviewers. Claude2 continues its existing feedback path and sends practical incoming-package findings when available; it does not acquire ownership of the independent review directory.
Existing authorization permits push and update of the Crown Key site. Keep R01 as dated history while the current pages clearly point to R02. Do not wait for another blanket publishing approval; ask Legend only for a specific unresolved creative/canon choice that the work actually needs.
Checkpoints and the rolling buffer
These are coordination targets measured from each seat's start, not promises of finished buildings at a fixed hour.
- First 30 minutes: return base SHA, exact claims, source list and first blocker. CLAUDE2 follows the existing Quik brief and records the actual Crown starting point and expected next handoff checkpoint without inventing an ETA.
- First 60–90 minutes: OA1 presents the atlas coverage/gap list; OA2 presents the bridge/topology resolution and route/room program; OA3 presents the Terrace room/entrance plan; Claude1 presents the shortest existing-pipeline path and blocked inputs.
- Every subsequent 60–90 minutes: one integrated checkpoint per asset with actual files, thumbnail sheet/drawing, changed decisions and remaining blockers. Prefer one coherent increment over a large unreviewed batch.
- Before Quik reaches his agreed Crown handoff checkpoint: target Split build-ready, Terrace independently reviewed and converging on the second ready slot. If Split is blocked, release Terrace first when it meets the same bar. Report the changed order explicitly.
- After each build-ready issue: its owner starts the next family; the other owner finishes the second queued package. Replenish the buffer before starting lower-priority image variants. If the buffer reaches zero, concentrate work on the nearest releasable packet.
The status board lists building / owner / state / evidence link / blocker / next deliverable, plus ready buffer = 0, 1 or 2+. Measure cycle time from the first completed packet and adjust work-in-progress limits from evidence. No fabricated throughput guarantee.
First launch
Open AI_START_HERE.md. Launch OA2, OA3 and CLAUDE1 with their matching briefs under seats/. Give Quik's already assigned Claude the CLAUDE2 coordination addendum without restarting its work. OA1 starts atlas/scope/site reconciliation and arranges independent read-only review while the two asset leads develop Split and Terrace, Claude1 prepares deterministic delivery, and Quik's Claude continues the existing fixture/sample/Crown execution. No further general questions are needed to begin.
