Image to HTML / CASE 01 / 2026-08-16
One reference image, two Agent Harnesses: what can each one build?
Gemini 3.7 Flash runs in Pi Harness and in DeepSeek Harness with the same reference image and the same task, generating an interactive ramen ordering web page in each.
- Model
- Gemini 3.7 Flash
- Input
- Same reference image
- Prompt
- Same
- Variable
- Agent Harness
EXPERIMENT SETUP / 01
Shared task
Starting from the same ramen ordering reference design and the same Prompt, generate a responsive, interactive ordering page. The two runs change only the Agent Harness; the model and the input stay the same.
The material pack records that both runs used the same Prompt, but it does not include a separately saved verbatim Prompt that can be shown publicly, so this page does not reconstruct or guess the original text.
RESULTS / 02
Two runs, two complete products
Clicking a result opens the full HTML in a new window, without the site shell around it, so you can use search, customization, the cart and the checkout flow directly.
RecommendedRUN 01 / Pi Harness
YAMATO RAMEN
- Time
- 22 min
- Cost
- $0.38
- Score
- 73/100
StrengthMore mature information architecture, filtering, depth of customization and overall brand feel.
LimitationThe mobile header is crowded, and the checkout form and default tip still need fixing.
RUN 02 / DeepSeek Harness
RAMENYA
- Time
- 28 min
- Cost
- $1.18
- Score
- 70/100
StrengthA more complete purchase loop, mobile drawer, order tracking and status feedback.
LimitationThe visual direction drifts from the reference, the page is too long, and the product image problems are more obvious.
COMPARISON / 03
Time, cost and results
Scores come from a combined check of actual interaction, responsive behavior, visual fidelity, information architecture and engineering completeness.
| Harness | Model | Time | Cost | Score | Conclusion |
|---|---|---|---|---|---|
| Pi Harness | Gemini 3.7 Flash | 22 min | $0.38 | 73/100 | Main branch for further iteration |
| DeepSeek Harness | Gemini 3.7 Flash | 28 min | $1.18 | 70/100 | Adopt some of its capabilities |
VERDICT / 04
Pi Harness wins, but neither run is a finished product yet
The Pi Harness version wins with 73 points: its information architecture, search and filtering, product customization and checkout summary are closer to something that could become a product. The DeepSeek Harness version is more complete in mobile navigation, order tracking and status feedback.
- 01
Both runs completed search, filtering, product customization, a cart and a simulated checkout, which already reaches the level of an interactive high-fidelity prototype.
- 02
Both runs used cropped pieces of the reference UI screenshot as dish images, producing an obvious "web page inside a web page". This is a shared P0 issue.
- 03
Both runs turned the five separate task pages into one long dark single page, a clear departure from the information architecture and visual direction of the reference design.
- 04
For further iteration, the Pi Harness version should be the main branch, taking in the mobile drawer and persistent order tracking from the DeepSeek Harness version.
KEEP EXPLORING