Video workflow comparison: motion design and architectural reconstruction

Assessment date: 2026-09-09. Source baseline: 836abf58de0346dbfe7bd1b318400f128b1dbbc7.

Verdict: Oshun has substantial reusable components, but the demonstrated workflow is not established as an available, connected Eve capability. Blender execution is real and was exercised on this Mac during this assessment. After Effects requires a live backend implementation and connection. The cross-application flows, creative review, skill library, and generation systems provide a foundation; their existence does not establish end-to-end availability or the visual quality demonstrated in the video.

This is an assessment, not an implementation or release-admission change.

What the video establishes#

Primary source: Bad Decisions Studio, GPT-6 just made THIS Motion Graphics in 30 MINUTES.

Segment	Reported process and limits
02:48–04:55	They describe 48 hours of reference-led correction, then a reusable motion-design skill and computer use inside After Effects.
10:58–15:17	Results involved roughly seven or eight versions, paid templates, tutorials, and corrections to glow, translucency, typography, and motion.
18:35–21:44	For an original sequence: generate storyboard alternatives and assets, compare them in an HTML dashboard, refine the selected storyboard, then produce an editable After Effects project.
24:22–26:38	Read listing photos and floor-plan dimensions, model and light in Blender, then import into Unreal. They explicitly describe an unfinished reconstruction with possible measurement/design differences.

Interpretation: “training” here describes instruction, feedback, and skill authoring; the evidence does not establish model-weight fine-tuning. The title does not establish a 30-minute end-to-end production budget. Exact private prompts, skill contents, and the application's click-versus-script execution mix remain unverified.

Comparison against current Oshun code#

“Implemented component” means source logic exists at the named boundary. “Live component” means an actual application was exercised. Neither label means the complete user journey is available through Eve.

Required capability	Existing Oshun system and evidence	Assessment
Gather web and visual references	Assistant built-in tools include web fetch/search and optional browser, vision, filesystem, and image-generation tools.	Implemented components. Powerful tools are mounted only when configured; this is not proof they are exposed together on the intended Eve surface.
Preserve a learned creative procedure	Skill tools implement save/search/run; skill seeding converts templates, workflows, and usage patterns into recipes.	A strong conceptual match. No benchmarked After Effects motion-design recipe was identified in the examined skill implementation. A stored recipe cannot supply missing tools or establish design quality.
Plan shots and generate assets	Isis storyboard-to-shot compiler handles frame order, duration, inherited parameters, and continuity. Assistant media tools expose provider-backed image generation.	Useful implemented components. Not a verified reference-to-editable-AE composition pipeline. Provider readiness was not tested in this assessment.
Execute a creative production plan	Yemaya's film pipeline factory builds dependent research, asset, DCC, rendering, post-production, and export steps.	Existing orchestration structure. Building a pipeline configuration does not prove its runtime dependencies or artistic result.
Author in After Effects	Bellona's Adobe adapter exposes `process.aftereffects.comp.read`, `property.set`, and `script.execute`.	Contract and dispatch logic exist. The shipped unconfigured backend reports unavailable; the examined production call sites do not construct a live AE backend. No dedicated AE render/save/export command is declared here.
Create and export editable Blender scenes	Blender RPC bridge executes real Python/operators through stdin or WebSocket; the remote-host adapter has asset and rendering command families.	Live component: direct stdin creation, save, reopen, semantic readback, GLB and OBJ export passed here. This did not exercise Eve, the remote-host WebSocket path, or artistic scene quality.
Transfer Blender assets to Unreal	Creative flow orders export, transfer, import, placement, screenshot verification, and provenance. Unreal adapter includes later import and editor-transport surfaces.	Implemented orchestration with injected providers. Tests use fake application providers. The older flow's comments about missing stage-23 commands must not be interpreted as proof those commands are absent everywhere today; the unresolved claim is live wiring and verification of the composed flow.
Review alternatives and request corrections	Yemaya dailies review models frame notes, annotations, comparison modes, and decisions. Unreal-to-Yemaya collects artifacts and attaches them to review.	Review primitives and a composed flow exist. No tested closed loop from a motion-design note to a corrected native AE project was established.
Retain editable sources and lineage	Creative-flow providers, artifact storage, and the asset dependency graph thread hashes and source/output relationships.	Existing foundation for project versions and review artifacts; completion must include reopenable source projects as well as preview media.
Execute desktop work autonomously through Eve	Browser/native ownership ADR, native fixture ADR, and current Phase 6 ledger.	Native execution/control evidence exists on Linux X11. Task 6.6 admission remains open; it does not establish macOS After Effects operation or availability in the member drawer.

The decisive integration gap#

The September wiring audit identifies real Bellona components without a connected production Eve execution path. The following source checks still support that finding at the baseline above:

The shipped MCP CLI calls connectBellonaMcpGatewayStdioServer() without options.
The stdio server registers DCC remote-control tools only when remoteControlGateway is supplied.
The BFF's Bellona imports found in the inspected TypeScript source concern asset package building, not constructing this gateway and a live DCC bridge.
The current Eve work ledger still leaves 5.5–5.8 open for attributed, leased DCC execution and its end-to-end evidence; 5.10 and 5.13–5.14 cover governed Unreal delivery. Task 6.6 remains open for native computer-use admission even though 6.4 and 6.5 have narrow native fixture/control evidence.

The corrected note under 5.5 says the measured cross-phase Security gate passed. It retains unfinished dependency and runtime-wiring requirements. Repeating the older “Security gate failed” note alone would misstate the current ledger.

The existing ownership decision routes DCC execution through attributed leased work, with task, actor, authority, budgets, and artifact evidence. Connecting arbitrary desktop commands directly to the copilot drawer would bypass that design rather than complete it.

What can be supported now#

An engineering-led Blender proof of concept is feasible now on this Mac. The direct bridge can produce and export a real editable scene. A developer must still assemble the task-specific interpretation, execution, and review loop. This is narrower than an available Eve product workflow.

A matching After Effects workflow needs implementation, not just a setting. The existing adapter offers an appropriate extension point, but a live backend must actually connect to AE, inspect compositions, execute bounded work, save projects, obtain renders, and return verifiable results. A generic script.execute contract does not establish any of those operations on its own.

Accurate architectural reconstruction needs a separate acceptance model. No dedicated listing-to-dimensioned-house implementation was identified in the examined Bellona, Isis, and Yemaya production paths. It should preserve measured versus inferred dimensions, units, room adjacency, openings, floor heights, source references, and uncertainty. Mesh validity or an attractive screenshot does not establish those properties. Unreal import also needs scale, material, camera, collision, and interaction checks. A browser-deliverable walkthrough requires its own delivery path and acceptance test.

Recommended implementation sequence#

Reuse the existing responsibilities: Eve coordinates the request and reusable procedure; Isis generates reference assets; Bellona controls native creative applications; Yemaya owns project planning, review, and delivery.

flowchart LR A[Brief and references] --> B[Storyboard alternatives] B --> C[Human selects and annotates] C --> D[Assets and reusable procedure] D --> E[Attributed Bellona application task] E --> F[Native project and rendered preview] F --> G[Readback and creative review] G -->|Revise| E G -->|Accept| H[Versioned project and deliverables]

Close the existing Eve-to-Bellona work boundary. Complete the retained 5.5 dependencies, bind the gateway to the attributed task, and demonstrate execution, cancellation, recovery, and independent output verification.
Prove one small Blender job through that boundary. Use a parameterized room or product scene with known dimensions, one requested revision, native save/reopen, and exported artifacts. The direct probe below is a starting point, not acceptance of this stage.
Connect the AE adapter on an appropriate host. Use a short typography sequence with layered shapes, easing, glow, and a camera move. Require a reopenable project, a preview render, and readback of layer/keyframe facts.
Close the storyboard/review/revision loop. Carry frame-specific notes and approved assets into the next execution; retain each version and its render. Save the successful recipe only after repeating it on a different brief.
Prove the house workflow separately. Start with a small measured plan, then test Blender-to-Unreal import, known dimensional tolerances, visual comparison, and interaction. Add listing ingestion after the geometry and verification path works.

Evaluate creative quality with an explicit rubric: typography, spacing, hierarchy, timing/easing, shape and path fidelity, materials, lighting, and temporal artifacts. Existing image/mesh metrics are useful checks, but cannot replace this judgement. Track elapsed time and human interventions across the entire task, including references, revisions, and review.

Verification performed during this assessment#

Host: macOS arm64, 16 GiB RAM, approximately 3 GiB configured swap and 10 GiB disk available. Other agent/CPU work was active. Resource inspection preceded single-worker, targeted commands; no broad Nx/build sweep was run.

Check	Result	Proving boundary
Adobe adapter tests	16 passed	Contracts, dispatch, validation, refusal, and rollback against an in-memory backend; no Adobe application.
Isis-to-Blender, Blender-to-Unreal, Unreal-to-Yemaya tests	14 passed	Flow ordering, artifacts, unavailable providers, and failures; fake DCC providers.
Assistant skill tools, skill seeding, media-tool tests	34 passed	Reusable recipe and provider-interface behavior; no live model or image-provider call.
Existing live Blender executable smoke	1 passed, 1 failed	Missing-executable refusal passed. Live save reached a header assertion that expected `BLENDER`; this installation wrote a compressed file. Export assertions were not reached in this run.
Supplemental direct-bridge probe	Passed	Blender 5.0.1 created `WorkflowStudyCube`, saved the default file, reset the scene, reopened the saved file, verified location `[1,2,3]` and dimensions `[2,2,2]`, then exported GLB and OBJ.

Reproduction commands, run from their respective package directories:

# libs/bellona/remote-adapters
pnpm exec vitest run src/photoshop-aftereffects-adapter.test.ts --maxWorkers=1

# libs/bellona/creative-flows
pnpm exec vitest run src/blender-to-unreal.test.ts src/isis-to-blender.test.ts src/unreal-to-yemaya.test.ts --maxWorkers=1

# libs/oshun/assistant
pnpm exec vitest run src/skills/skill-tools.test.ts src/skills/skill-seeding.test.ts src/tools/media.tool.test.ts --maxWorkers=1

# libs/bellona/blender-agent: retains the observed failing header assertion
BELLONA_AUDIT_BLENDER_PATH=/Applications/Blender.app/Contents/MacOS/Blender pnpm exec vitest run src/blender-executable-smoke.integration.spec.ts --maxWorkers=1

The supplemental probe used the shipped createBellonaBlenderRpcBridge with stdin, background: true, and factoryStartup: true. Its script called primitive_cube_add, save_as_mainfile, read_factory_settings, open_mainfile, export_scene.gltf, and wm.obj_export. Node assertions checked the reopened object facts, GLB magic, and OBJ vertex/face records. The scratch directory and Blender process were cleaned up. Observed artifacts:

.blend: 92,998 bytes; first eight bytes 28b52ffd60190535; SHA-256 de527011d9707b61174eb590e0fd99fff08be25a86961704d748d9ec620b12db.
.glb: 3,480 bytes, glTF magic.
.obj: 1,870 bytes, vertex and face records present.

These sizes and hash identify this probe, not expected values for future runs. The existing smoke needs a compression-aware assertion or semantic reopen verification. Its failure is retained as an assessment finding; no test or runtime code was changed to make the result green.

The mandated Linux Unreal path was checked and was absent on this Mac. No After Effects installation or Unreal installation was found in the application locations inspected. The earlier Linux host matrix records UE 5.5.4 on another host; that is not current Mac availability evidence. No live UE, AE, image-provider, or complete Eve workflow was tested here.

Evidence limitations#

The video analysis used its publisher description, auto-generated English transcript, and sampled YouTube storyboard images showing the AE and Blender/Unreal sequences. It was not a full-resolution frame-by-frame review. The in-app browser execution tool was unavailable in this session; public metadata, captions, and storyboard images were retrieved directly instead. Private course materials and unpublished skill files were not accessed.

Downloaded transcript/images and temporary probe logs were kept outside the repository. This report contains the analysis and verification results. Its passing component tests do not erase the failed existing smoke or establish production readiness. Product UI, runtime behavior, schemas, and release gates were not changed; no UI or application build verification is claimed.