Checked the camera before using credits — Previs made with Claude Opus 5.5 in three.js, 2 short f...

Summary This video by the Korean tech channel AgentOS demonstrates how to pre-visualize AI video scenes using three.js 3D HTML files generated by Claude Opus 5.5 before spending generation credits. By connecting Claude to Higgsfield via Model Context Protocol (MCP) and generating rough blockings, camera angles, and timings with simple geometric boxes, the creator directs full-fidelity videos in Higgsfield’s Seedance 2.5 model across martial arts and horror genres.

---

What is shown

• Side-by-side comparison [00:00–00:15]: Comparing a simple 3D box previz in an HTML file against the final AI-generated video in Seedance 2.5, matching camera trajectory, cut timing, and character position.
• Claude Opus 5.5 Three.js Previs Generation [01:43–02:35]: Prompting Claude Opus 5.5 with a single-sentence brief ("rainy bamboo forest, female swordfighter fighting multiple assassins in a 20s scene") to code a standalone three.js HTML file containing 20 choreographed moves, 13 camera cuts, a top-down radar map, and playback scrubbing.
• Previs Features [02:35–02:57]: Subtitle timing lanes, voice pre-listen options, playback speed reduction (0.25x), and conversational camera adjustments in natural language without using credits.
• Higgsfield MCP Integration [03:25–04:03]: Connecting the MCP server (https://mcp.higgsfield.ai/mcp) into Claude's connector settings, generating a character turnaround sheet with a blanked face on the full body to prevent model confusion, and submitting 2x 10-second prompt batches to Seedance 2.5.
• Completed Bamboo Forest Action Scene [04:04–04:23]: The final cinematic wuxia clip rendered by Seedance 2.5.
• Backrooms Time-Loop Test [04:24–05:29]: A 30-second psychological horror scene in yellow corridors showing how tilting box angles translates to leaning around corners; includes playback of the final generated film.
• Higgsfield Motion & VFX Ecosystem [05:30–06:40]: Demonstrations of Opus 5.5 driving After Effects composition scripts, Blender building demolition simulations, the 11-skill "Production Skills Bundle" (/destruction-studio, /shot-cleanup, /shot-composer), and analysis of community project Red Flag.
• Case Study: Chloe VS History [06:45–07:11]: Examining a historical travel-vlog AI channel created in August 2025.

---

Claims & numbers

• Model releases & pricing (presenter citing Anthropic & Higgsfield):
  • Claude Opus 5.5 was released the previous week (September 22, 2026), priced at $4 / million tokens input and $20 / million tokens output (prompt caching: $0.20 read / $5 write).
  • Opus 5.5 represents a 20% price cut compared to Opus 5 ($5 / $25) and costs 40% of flagship Claude Fable 5.1 ($10 / $50).
• Generation costs & specs:
  • Seedance 2.5 charges 70 credits per 10-second 720p generation.
  • A 20-second wuxia scene generated in two concurrent 10-second segments took approximately 5–6 minutes per segment and consumed 140 credits in total (graphic shows 133–140 credits).
  • Generating the initial three.js HTML previs file with Opus 5.5 took 12 minutes of model computation.
• Backrooms previz timing test:
  • Dialogue and light flickers lagged the HTML previs keyframes by approximately 0.4 to 1.4 seconds in the final output, though shot sizes, angles, and cut orders aligned precisely.
• Higgsfield Production Skills Bundle:
  • Includes 11 standardized creative pipeline skills for Blender, After Effects, and TouchDesigner.
• Community statistics:
  • Red Flag (Higgsfield community short): 2-minute runtime created without physical sets or cameras, utilizing 3,528 generations and over 3,500 assets.
  • Chloe VS History YouTube channel: Registered August 4, 2025; has over 406,000 subscribers, 59 videos, and over 29.28 million lifetime views (with its Titanic video reaching 2.81 million views).

---

Notable quotes

• [02:56]: "여기까지는 크레딧이 하나도 들지 않아요." (Up to this point, not a single credit is spent.)
• [06:36]: "구도는 다이어그램이 정하고, 장소 사진은 질감과 빛만 맡는다." (The diagram decides the frame; the location photo only handles texture and light.) — quoting the production notes of Higgsfield's Red Flag.
• [07:16]: "결국 본인의 관심사와 취향이에요." (In the end, it all comes down to your personal interests and taste.)

---

Assessment This is a high-quality video essay and technical workflow tutorial by a creator demonstrating hands-on generative AI filmmaking techniques. The demonstrated tools (Claude Opus 5.5, three.js HTML generation, Higgsfield MCP, and Seedance 2.5) are shown functioning in real software interfaces, candidly highlighting practical nuances such as a 0.4–1.4 second timing drift between previs blockings and video model generations.

---

Lyrics & themes

• Structure & Narration Themes:
  • 00:00–01:42 (Cost & Previs Rationale): Highlighting the dilemma of expensive generation credits and proposing zero-cost browser-based three.js previz over bulky 3D software like Blender.
  • 01:43–03:24 (Prompting Code for Spatial Control): Walking through prompt mechanics to produce interactive 3D choreography timelines.
  • 03:25–05:29 (MCP Execution & Validation): Connecting the workflow directly to video generators and verifying timing accuracy across martial arts and Backrooms horror.
  • 05:30–07:38 (Democratization of Directing): Discussing automated VFX scripting and emphasizing that as technical barriers dissolve, personal directorial taste becomes the key differentiator.

---

Lore & references

• Backrooms & Yellow Corridors [00:00, 04:28]: The creepypasta concept of the "Backrooms" is used as a benchmark scene to demonstrate spatial orientation, duplicate character encounters, and time-loop consistency.
• Higgsfield MCP: Reflects Anthropic's Model Context Protocol standard adopted by third-party generative media platforms to let Claude directly call image/video APIs.
• AI Cinema Discourse (Red Flag & Chloe VS History): References early milestone AI productions that prove entire film sets, historic reenactments, and episodic YouTube travel channels can be executed without physical cameras or crews.

---

Visual style & craft

• Visual Presentation: Structured educational presentation featuring clean typography, split-screen video comparisons (three.js canvas on top, rendered AI footage on bottom), screen recordings of browser sessions, and dark-mode infographics.
• Production Attribution: The tutorial itself is cleanly edited with human pacing, diagrams, and motion graphics, while the featured film sequences are AI-generated by Seedance 2.5 guided by Opus 5.5-written three.js code.

Described by gemini-3.8-flash on 2026-09-29 from the video's audio and frames.