Where AI speeds up my design work — and where I keep the responsibility

From fragmented inputs to a designable question

In Corevia, AI helped me structure four roles around a shared order lifecycle; in ParentMind Hub, it helped organise desk research, JTBD, and flows around SOS and Recovery. I ask it to expose conflicts between shared events and role-specific decisions, and between survey signals and untested conclusions.

My rule: bring verifiable material, mark assumptions, and ask for failure paths and open questions. A brief is not a field study. Thirty-one parent responses cannot be converted by AI into proof of intervention effectiveness.

• Input: briefs, screens, documents, open questions.
• Output: IA, flows, options, and validation checklists.
• Review: return to the source and my actual contribution.

ChatGPT helps me build and audit, not decide for me

Across portfolio and UI artifacts, I have used ChatGPT to build reviewable prototypes and pages, write EN/VI copy for each context, and inspect component, responsive, and interaction inconsistencies. I have also had to catch missing Auto Layout, lookalike components without instances, unstyled text, and mismatched icons. Speed does not guarantee quality.

I treat every generated screen as a proposal to QA: are labels understandable, states meaningful, Vietnamese diacritics visible, and failure paths recoverable? If visual QA or post-launch data is missing, I say so instead of declaring an unverified improvement.

Claude in my workflow: a second lens, not a fabricated achievement

Claude is among the AI tools I am exploring. I prepared a brief for Claude Extension to review the structure and starter content of UX Learn & Share on Discord, and looked into Claude ↔ Figma MCP. Preparation does not mean Claude completed that community or built a Corevia Figma screen.

The method I am developing is to give one tool a sourced brief to structure options and use another as a reviewer of assumptions, edge cases, and UX writing. I keep only what I can explain from the product context and artifact. This is an evolving practice, not a validated automation or usability result.

The boundaries I will not outsource

AI cannot interview parents, review sensitive guidance as a qualified expert, observe a driver in weak connectivity, or measure post-launch business impact. Personal data should not be poured into a prompt for convenience.

If AI produces ten plausible solutions, my job is to choose the right question, discard unsupported claims, articulate trade-offs, and validate with people. The goal is a better decision—not simply a faster screen.

• Separate project facts from hypotheses and AI suggestions.
• Review edge cases, privacy, and implementability.
• Never claim a test or outcome that did not happen.