A running recap of our June 10–11 sessions — what you set out to build, the questions you raised, and what I'd recommend next — so nothing gets lost between now and your build. Push back on anything here. — Ben
▶ Start here
Nader builds the PRD — let Claude interview you on the workflow, then we review it together. That one step unblocks everything else.
What you wanted to build
The goal: turn a 2D floor plan into an accurate, web-ready 3D render for sales & marketing — built in reliable, checkable steps, not one risky prompt.
- ◆A web app that reads a 2D plan into structured data, with editable fields to correct it.Live
- ◆Compare AI models in the app to weigh accuracy vs. cost.Live
- ◆Accuracy scoring — grade the AI's read against a human-entered "truth."Live
- ◆Prompt Studio — an internal tool to tune the prompt, add memories & knowledge, push accuracy.Mockup
- ◆Split multi-floor sheets into one floor per image for sharper reads.Next
- ◆PDF input — most builder plans arrive as PDFs, not images.Next
- ◆A tuning loop — the AI keeps adjusting until its output matches your "truth."Next
- ◆Web-ready 3D renders from the validated specs, via a multi-step pipeline (model → detail → styling → QA).Later
- ◆Targeted edits ("fix the stairs," "change the wall colour") instead of regenerating the whole image.Later
- ◆AI virtual staging — room photo → pick a style → staged image (possible second app).Idea
Decisions & next steps
- 1Phase 1 = extract + validate only. No 3D rendering until extraction is reliably accurate.
- 2GPT-5.5 chosen as the extraction model (most accurate in testing); next is tuning against human-entered "truth."
- 3Move into Empire's Cloudflare account — targeting early the week of June 15.
- 4Back to planning: build a PRD (let Claude interview Nader), then review it with Ben.
- 5Team starts building the week of June 15, once the environment move is done.
- 6Ben to share both session recordings + his development playbook; Prompt Studio mockup is live (links above).
Open questions to settle
- ?Priority: 3D floor-plan generation vs. virtual staging — which comes first, and how do we split effort?
- ?Multi-floor handling: confirm the split-into-single-floor approach before building it.
- ?Validation owner: who on the team signs off that the AI's specs are correct?
- ?"Good enough" planning bar — how much PRD detail before we start building, so we don't stall?
- ?Gemini: revisit once its API is stable, or leave it out of the lineup for now?
Ben's recommendations
- ✓Use GPT-5.5 for extraction now — most accurate in our testing (25 rooms vs. Claude's 23, with fewer uncertainties). Claude is better for code.
- ✓Separate every concern — don't ask the AI to model 3D, style it, and lay out a branded page in one step. Each step has to succeed before the next begins.
- ✓Plan before you code — write a short PRD first (dictate it to Claude). A sharper plan means a faster, cheaper build.
- ✓Accuracy comes from a pipeline, not one prompt — extract → review → QA, not a single magic call.
- ✓Human-in-the-loop is good process, not a failure — expect a team member to confirm specs before any 3D render.
- ✓Prompt-tune, don't "train" — build a knowledge base of your symbols and terms; grow it as you find gaps.
- ✓Your team runs this day-to-day — the plan is for your people to own the workflow, not route everything through ZeroArc.
Questions you asked → answers (reference)
- How does the dashboard know how accurate it is vs. the truth?→ You enter the correct specs once. The system compares the AI's output to your numbers and scores the gap.
- Where does that "truth" get entered?→ Right on the review screen — an editable panel where a person confirms or corrects each field before the comparison runs.
- Are we permanently "training" the model so it learns over time?→ No — real training costs billions. You tune the prompt and build memories; the model doesn't change, your instructions do.
- Does the model know it's looking at three separate floors, not one house?→ Not unless you tell it. One added hint and the read improved immediately — hence splitting multi-floor sheets.
- Can we teach it our symbols — stairs, wall types, WIC, etc.?→ Yes. That's prompt-tuning: build a knowledge base of your symbols and terminology and feed it to the prompt.
- Do we need an architect or designer to verify the specs?→ Probably not — the fields are plain-English (rooms, dimensions, connections); a team member can check them.
- Do we re-enter the truth for every single plan?→ No — tune on a representative sample, then it generalizes to similar plans; after that it's a quick spot-check.
- Why do the AI images get worse the more we iterate?→ The model rebuilds the whole image from scratch each time — there's no real "edit," so good parts get lost.
- Unified billing vs. bring-your-own API keys?→ Unified billing — one place, every model, fully logged. BYOK is only worth it for tighter access control.
Notes by Ben (ZeroArc). Push back on anything here and we'll adjust.