*Sensitive information has been removed or redacted in accordance with non-disclosure agreements.
Snapshot
About: AI-guided product discovery that helps an individual user turn an early-stage agent or application concept into a tested, implementation-ready specification through structured discovery, ideation, workflow design, usability evaluation, and iteration.
Role & scope: Creator, UX strategist, workflow designer. Designed and built the end-to-end AI-orchestrated product discovery system.
Team: Independent project
Timeframe: 3 months
Status: Functional system completed; seeking pilot partner.
Key Decisions (type / authority):
Scope: Decided
Workflow structure: Decided
UX Methodology Logic: Decided
Orchestration Architecture: Decided
Implementation Spec Structure: Consulted with developers
Problem
Opportunities I noticed:
AI has democratized application development: A single user can quickly turn an idea into a prototype or working agent without a traditional product team.
That speed creates a design quality gap: Internal stakeholders often jump to solutions before fully articulating the underlying problem, while much of the workflow knowledge needed to design an effective system remains tacit and never makes it into the prompt.
AI integration is fundamentally a workflow-design problem: The challenge is not just what the model can do, but how responsibilities should be divided across humans, deterministic systems, and AI according to their respective strengths.
Reframed problem: How might we help individual builders turn early-stage AI ideas into well-designed, implementation-ready specs?
Options and Tradeoffs
Speed vs Depth: Too slow and the process becomes a hurdle; too shallow and it loses methodological value.
Decision: Preserve only the non-negotiable steps needed for rigorous problem framing, workflow design, and evaluation.
AI vs Human Autonomy: Too much AI control can override the user’s context and judgment; too little reduces its ability to challenge assumptions, reframe narrow asks, and surface broader options.
Decision: AI steers the methodology, while users provide context and retain decision authority at key points.
Monolithic workflow vs modular skills: A hard-coded flow is faster and easier to control; modular skills are more reusable and extensible but add complexity.
Decision: Ship the MVP as a controlled orchestrated workflow and modularize later.
Design
Method Development:
Explicit Knowledge: Reviewed UX research and design references, contributed my own best practices, and synthesized common principles and techniques into a coherent methodology.
Grounded-theory analysis methods
Service-design and workflow-design principles
Human factors / human–AI function-allocation principles
Lessons from AI-assisted synthesis workflow
Tacit Knowledge: Worked through sample data and training use cases in parallel with the system, comparing its approach with my own. Gaps between our outputs surfaced principles and heuristics I had relied on implicitly but had never articulated.
Objectives
Help the user identify the problem behind their initial ask. Surface tacit context and look for the leverage point upstream rather than accepting premature solution framing.
Help the user allocate work intentionally. Assign responsibilities across humans, deterministic systems, and AI based on their strengths, not on what’s convenient to delegate.
Help the user validate workflow logic before implementation. Run the specification as a scenario to expose missing steps, sequencing failures, and unclear handoffs that may look correct on paper.
Challenges
Creating conditions for real disclosure: Tacit knowledge often surfaces through the rapport, pacing, mirroring, probing, and judgment of a skilled moderator. The challenge was designing an AI interaction that could approximate this without a human present.
Preserving depth in text-only medium: Text lacks the nonverbal richness and fluid back-and-forth of voice, so questions had to be more precise to preserve depth without creating friction.
Encoding interpretive judgment: Research analysis depends on subtle judgment: reading ambiguity, contradictions, emphasis, and context. Translating that practitioner intuition into explicit rules and heuristics was difficult.
Evaluation
Method: I evaluated the system myself, running realistic use cases through the full workflow. No external users were involved.
Findings
Insufficient interview-stage data is a primary constraint. Conversation pacing during the interview stage balances completeness of data against the user’s expectation for speed and efficiency. When the interview under-collects, the shortfall cascades downstream, surfacing as workflow inaccuracies during the usability testing phase.
Problem reframing worked, conditional on interview depth: Given sufficient interview data, the system applied an interpretive lens and produced a useful reframing of the underlying problem structure. It is one of the strongest aspects of the system.
Ideation lacks momentum compared to group brainstorms: No other peers contributing to lower the stakes of a bad idea, no human facilitator helping the participant feel comfortable getting creative, and none of the spark that makes group brainstorming feel exhilarating rather than cerebral. Even with the system actively guiding the process, momentum still falls entirely on the user.
Cross-stage continuity was strong: Context handoff remained coherent across every stage of the pipeline.
Limitations
The system has been completed and exercised as a functional workflow, but it has not yet been evaluated with external users on live projects.
Behavioral analysis: The system does not have the capability to observe user behavior directly, so all workflow and behavioral insights are inferred from conversation. There would need to be an external source of behavioral data or future capability to analyze a video of users interacting with a prototype.
Whats Next
Actively seeking a team or project to pilot this system against a real early-stage concept.
Move from hard-coded workflow to modular skills.
Find a valid and reliable way to extract behavioral insights.
Facilitate research with multiple distributed users to feed a centralized project knowledge base.
Longer-term inquiry: The current tool is based on best practices developed for human UX practitioners. With AI mediating the process, how should those methods be reconfigured rather than simply replicated?