Porting an Opus 5.5 motion design workflow to Gemini 4 Argon by renaming CLAUDE.md breaks in production, author says

Emily @IamEmily2050Originally posted 2026.10.11

Open original on X

So much happening at the moment especially on motion design thanks to Opus 5.5 and because Gemini 4 Argon around the corner I would like to share my though about the situation. Porting a Claude Opus 5.5 motion graphics workflow directly to Gemini 4 Argon by renaming CLAUDE. md to GEMINI. md breaks in production because the two models have opposite architectural strengths and bottlenecks. 1. Interactive Shell Loops vs. Single-Pass Repo Synthesis Claude Opus 5.5 leads on interactive terminal tasks (66.4% vs. Argon's 57.4% on Terminal-Bench 4.0). Claude works best when making small, iterative bash edits shot by shot. Gemini 4 Argon leads on full repository architecture (77.9% on DeepSWE v1.1 and 91.9% on Vibe Code Bench) and ships with a 1,000,000-token output ceiling. Running Argon in a chatty trial-and-error shell loop plays straight into its weakest benchmark. Instead, push 90% of the work into a single uninterrupted generation pass where Argon outputs the physics library, dense SVG Bezier tables, procedural WAV synthesizer, and all scene files at once, followed by a single automated QA command (node checks/qa.mjs). 2. Subagent Stitching vs. Continuous Multi Scene State In code rendered video, a single 12 second scene can burn over 30,000 output tokens once internal reasoning is included. On smaller output windows, multi scene videos have to be split across subagents, and the seams frequently break because camera coordinates and spring momentum reset between files. Argon's 1M output window holds an entire 45 second, 10 scene repository in a single trajectory, allowing Scene A to hand off its exact analytical exit position and velocity directly into Scene B's entry spring. 3. Static 0.5s Contact Sheets vs. Native Video & Event Strip QA Opus 5.5 only takes text and static images. Reviewing motion requires slicing the rendered MP4 into 0.5 second PNG contact sheets with ffmpeg. At 1 frame every 0.5 seconds (every 15th frame at 30 fps), a 2 frame letter stagger or a 4 frame impact flash falls completely between the screenshots, so the model grades animations it never actually saw. Argon is built for native long video comprehension (91.7% on LVBench) and can inspect the rendered MP4 directly for optical stutter, bad spring damping, and 1 frame audio drift. When running on image only API endpoints, pair the 0.5s overview sheet with 1 frame interval event strips across every cut and flash. 4. CLI Plan Gates & Token Economics To keep costs down and prevent premature file edits: Enforce Gate 1 with Gemini CLI's read only /plan mode (gemini approval mode=plan) so the agent inspects your local assets and writes style_guide.md and shotlist. md before touching code. Pin GEMINI. md, beats.json, and local reference assets in a static prefix at the top of the prompt to trigger Argon's cached input discount across all review rounds. When a critique round fails, revert only the code (git checkout HEAD~1 -- src/ index.html) while keeping docs/review_log.md intact so the agent does not wipe its own memory and repeat the same broken fix.