AI Stack Rationale
| Role | Tool | Rationale |
|---|---|---|
| Workspace | Claude (Project) | An environment built for drafting, editing, and holding project context (brief, personas, prior decisions) across a long session without re-explaining background each time. I am also very interested in adopting the MCP workflow between Figma and Claude Code, then virtually all project context can live in Claude. |
| Research | Perplexity | Every claim comes with a numbered footnote linking directly to the original web page, making it easy to verify facts, cross-check information, and avoid black-box guessing. It performs dozens of searches, reads hundreds of sources, and reasons through the material to deliver a comprehensive report, typically completing in a few minutes. |
| Synthesis | Claude | All project files and work is contained within a Claude project so synthesis within Claude makes the most sense. It’s also very good at taking raw data or user interviews and extracting patterns, clustering ideas and turning them into structured artifacts (hypotheses, briefs) while flagging unverified claims, rather than treating AI output as finished fact. |
| Images | ChatGPT Image Creator and Firefly | ChatGPT used for fast, disposable exploration where volume and speed matter more than rights clearance. Firefly is where final assets get created because of its commercial-licensing positioning. |
| Prototyping | Figma Make | Fastest path to clickable prototype with interaction logic for usability testing. Because it operates inside Figma, prototypes stay connected to the design file’s components and styling rather than living in a separate tool. This shortens the loop between concept and testable prototype. Currently also experimenting with Figma MCP and Claude Code. |
| Communication | Gamma | Generates presentations, documents, and simple websites directly from prompts or pasted content. It has strong prompt-to-polish visual quality, great for fast stakeholder updates. |
My Working Pattern
- Frame the Project. Create the project brief, audience hypotheses, success metrics, and what would disprove each hypothesis.
- Research and Interviews. Run a market scan and competitive analysis with Perplexity and load findings and source validation log into project files. Conduct a series of interviews and load the transcripts into the project files.
- Synthesize. Use Claude to cluster patterns and themes, with all insights citing a source validation log ID. Separate stated preference from reported behavior, and identify the core user goals, pain points, and primary personas.
- Validate. Every claim audited against the primary source with sign off by me as the final step.
- Prototype. Create IA, sitemaps and user flow diagrams with Claude. Mood board creation with ChatGPT, initial concept screens in Figma then finally prototyped in Figma Make. Create a test plan and conduct usability testing. Log findings and load into Claude.
- Communicate. Create a stakeholder deck or case study in Gamma that clearly shows validated claims and additional hypotheses.
Ethics and Source-Validation Policy
Bias Auditing
Bias review happens at two checkpoints: immediately after any AI-assisted synthesis (personas, journey maps, copy), and again before stakeholder review. Each pass is checked against five criteria: age range assumed, body type/ability assumed, income or access to equipment assumed, culture or geography assumed, and motivation framed as universal (“everyone wants to feel their best”) rather than segment-specific. Anything that fails a criterion gets flagged and revised before it moves forward.
Source Validation
Every factual or demographic claim gets one of the four labels:
- Validated
- Confirmed via independent source/data
- Partially validated
- Supported but not confirmed at this scale/context
- Hypothesis
- Plausible but untested
- Discarded
- Investigated and ruled out
Disclosure of AI Assistance
Any deliverable with meaningful AI contribution states which tool was used and for what step (drafting, synthesis, image generation, prototyping). This disclosure appears in the deliverable itself, not just verbally in a meeting, so it travels with the file if it’s shared or reused later. If a stakeholder questions an AI-generated insight, it can be traced back to its validation label and source, showing the underlying evidence directly.
Organizational Policy
Every AI-generated asset that ships gets tagged with the tool used, prompt or method, and any human modifications.
My Current Prompt Library Structure
The library is divided into 6 high level categories that mirror my working pattern. Each high level category is then broken down into sub categories, so a prompt is found by where I currently am in my working pattern.
- 01Discover
- Research
- Users
- Problems
- Insights
- Personas
- Requirements
- 02Analyze
- Synthesis
- Patterns
- Findings
- 03Design
- Concepts
- Ideas
- Solutions
- UI
- Interaction
- Flows
- Content
- Prototype
- 04Test
- Usability
- Interviews
- Feedback
- Accessibility
- 05Handoff
- Stakeholder Decks
- Dev Specs
- Documentation
- 06Strategy
- Priorities
- Tradeoffs
- Roadmaps
What I’m keeping an eye on…
- Figma MCP write-to-canvas via Claude. If it builds screens with my tokens and components intact, move prototyping off Figma Make.
- The shift in design hiring expectations as AI competence becomes a baseline rather than a differentiator.
- The designer-to-developer boundary.