Meng To uses GPT-5.6 Sol Ultra to build an infinite React design canvas

• Meng To had GPT-5.6 Sol Ultra build an infinite HTML/React design canvas with layers, inspector panels, media items, and real-time collaboration, and Meng To's demo post says the model "pretty much one-shotted" it.
• The workflow was planner-first, worker-second: Meng To's workflow note says Sol breaks one goal into 20+ concrete steps, then each step becomes its own spawned thread.
• The demo had a scaffold: Meng To's context reply says context was key, while his codebase reply says the build sat on existing controls, Supabase, React/Vite, and HTML/React handling.
• The cost caveat surfaced immediately: Meng To's subagent reply says Ultra spawned enough subagents to be scary for tokens, and his pricing reply says token spend had already shaped his product pricing.

The prompt started as a messy voice note in a Universal Studios Japan line, according to Meng To's prompt reply. Meng later identified the environment as Codex browser use on localhost in his setup reply, while his Supabase reply made the realtime layer sound routine. The later pattern in Meng To's workflow note was cleaner than the prompt: one broad goal, one generated plan, then one spawned thread per task.

Meng asked Sol Ultra for an infinite canvas for HTML/React designs, with layers, an inspector, media items, and real-time collaboration. The useful part is not the Figma-shaped UI, it is how little repair work Meng said followed the generation.

dustinhollywood's reply called "no time fixing things" the real win. That is the line creative tool builders will remember.

The original ask was loose, repetitive, and specific in the right places. It asked for:

• a Figma-like infinite canvas
• multiple generated websites or React sites
• a left layers panel
• a right inspector
• HTML or React generation
• image and video items on the canvas
• multiple canvases
• proper planning before implementation

Meng later described his prompting style as increasing in complexity, in his prompt-complexity reply. The prompt reads like a product brief spoken before the ride line moves.

The build had home-field advantage. In Meng To's context reply, he said having context was key and suggested trying the pattern on an open-source project, an existing project with basic controls, or a project with lots of references.

Meng To's codebase reply named the inherited pieces: controls styling, backend, Supabase, React/Vite, and existing HTML/React management. The screenshot mattered because Sol was extending a design mode that already had similar interaction patterns, which Meng To's current-design-mode reply pointed back to.

Meng's second post gave the workflow more structure than the demo tweet:

• stay general while giving enough context
• ask open-ended questions
• give the agent the right skills
• keep the prompt pinned in an open Codex side browser, with a screenshot for context
• avoid being too vague or too specific
• ask the agent to make the plan
• let Sol split the goal into 20+ steps
• spawn one thread for each step
• commit after every change
• review each result independently
• roll back anything that fails

Codex is the harness around that pattern, and Meng To's Codex breakdown lists agent skills, goals, spawned threads, browser use, computer use, and mobile connector among the features he now uses to ship apps and organize content.

The model choice was layered too. Meng To's tool-switching reply says Ultra worked well for planning, xHigh for threads, and medium or high for smaller iterations.

The subagent swarm had a bill. Meng said Ultra spawned so many subagents that it was scary for his tokens, and his Codex-plan reply put his experimentation on a $200 Codex plan with resets while he figured out Ultra and xHigh.

Cost had already changed his product economics. Meng To's pricing reply says most tools had moved to token-based pricing, while his team stayed prompt-based after paying too much for tokens.

The broader GPT-5.6 week had the same pattern. Peter Yang said Sol High seemed to burn tokens faster than 5.5 High in his usage reply, and Dan Shipper joked in his overnight-token post that he accidentally spent 2b tokens overnight.

The run used Codex browser use against localhost. That detail matters because the agent could see and operate the app while building it, rather than generating code blind.

Realtime collaboration came from familiar infrastructure. Meng To's Supabase reply called it easy with Supabase, and his realtime reply reduced the implementation detail to: "Supabase has realtime."

Meng kept the output target plain. his HTML clarification says the build used "dead-simple html."

HTML was the interchange layer, not a downgrade. In Meng To's HTML reply, he listed landing pages, docs, slides, video through hyperframes, and conversion into React as reasons HTML stays useful.

dustinhollywood's reply put the workflow in one line: design in HTML, then convert.

Open-source entered the thread because the tool was suddenly cheap to create. In Meng To's decoupling reply, he said he was rethinking his open-source stance and would need to decouple the feature from an existing product.

Later, Meng To's open-source reply shortened the answer: he could probably decouple this and open-source it.