# Truvit — AI makes proposals. The designer makes the calls.

Shawn Du · Design engineering / AI creative workflow · 2026

Case study: https://shawndudesign.com/projects/truvit.html

## The problem

A designer developing a campaign needs more than an image generator. The workflow must carry source material, constraints and chosen direction into later revisions, while keeping assumptions and approval decisions visible.

## My design response

I structured Truvit around four decisions: confirm the understanding, choose a direction, accept the master and approve the current package. Six bounded capabilities support those decisions: context preparation, direction exploration, copy development, master composition, format adaptation and evaluation.

A workflow agent proposes a permitted revision route. A deterministic controller owns versions, permissions and execution. A separate reviewer call inspects the actual images and reports advisory findings. The designer remains responsible for creative judgment.

## The iteration that changed the project

When different briefs still felt visually similar, I challenged the output instead of treating successful model calls as success. Diagnosis found that two layout names generated identical files and some split-layout controls were ineffective.

The engine was expanded to five distinct compositions, three typographic voices, custom colors and format-specific geometry. Direction and master calls now receive the actual reference images. Direct edits create new files, preserve history and reopen review. The prototype composes supplied photographs; it does not generate product photography.

## What exists

A working private AI prototype, saved project history, editable campaign drafts, actual PNG and SVG outputs, and a 16-file technical export including fonts, copy and a manifest. The public case includes an edited 54-second walkthrough of real saved screens and files, so it can be assessed without access to the private app.

The cobalt technical run completed master, 4:5 and 9:16 output. The plum editorial example used live AI followed by directly edited copy and remains a draft awaiting owner creative approval. These are prototype demonstrations, not launched campaigns.

## Evidence and limits

Three participants completed the earlier fixed clickable prototype without prompts, according to my retrospective reports. Their feedback showed uneven comprehension, prompting an opening-copy revision that has not been retested.

The upgraded system has been exercised through real model calls, controlled rendering, revision checks and export verification. Those are agent-operated technical checks, not new participant research. No productivity gain, audience preference or commercial performance is claimed. The next study tests designers' comprehension of checkpoints and revision consequences.

## Contribution

I led product framing, workflow and interaction direction, iteration priorities and the original participant recruitment. Codex provided substantial implementation, debugging, technical testing and documentation support under my direction. The code was developed with AI assistance.

Independent portfolio prototype using supplied Moreno Bath example assets; no client commission or commercial relationship is claimed.