AI PRODUCT INNOVATION · PRODUCT STRATEGY · SYSTEMS DESIGN
Preflight: Designing AI to Think Before It Builds
An AI-powered reasoning system that helps product teams challenge assumptions, diagnose complex problems, and make more informed decisions before committing to solutions.
ROLE
Product Vision, Strategy, Experience Architecture, AI Development & Validation
PROJECT
Internal AI Product · Caylent · 2026
OVERVIEW
Making better decisions, not just faster ones.
AI has dramatically accelerated how quickly teams can generate concepts, requirements, designs, and code. But faster execution doesn't necessarily mean better decisions.
I conceived and built Preflight to address that gap: a structured AI reasoning system that helps teams investigate ambiguity, connect evidence, and determine what actually needs to change before deciding what to build.
The goal wasn't to replace human judgment. It was to make the reasoning behind it more visible, consistent, and actionable.
THE CHALLENGE
The solution was arriving before the understanding.
Product teams often jump from an observed symptom directly to an implementation idea. A confusing workflow becomes a UI redesign. Low adoption becomes a feature request. An AI opportunity becomes a chatbot.
But the visible symptom isn't always the underlying problem.
The challenge wasn't a shortage of frameworks or methodologies. It was the lack of a consistent way to connect evidence, uncover underlying mechanisms, and evaluate whether a proposed intervention would actually work.
FROM REACTIVE SOLUTIONS TO GROUNDED DECISIONS
WITHOUT PREFLIGHT
→
WITH PREFLIGHT
→
→
→
THE INSIGHT
AI didn't need another playbook. It needed a reasoning system.
A library of product frameworks would make knowledge more accessible, but accessibility alone doesn't improve judgment.
The real opportunity was designing AI to recognize what kind of reasoning a situation required, identify missing evidence, and connect problems to strategies through their underlying causes.
This shifted the concept from a methodology assistant to an adaptive product reasoning system.
THE SOLUTION
One intelligence system. Four ways to think.
I designed four specialized reasoning modes, connected through a Product Router that identifies the earliest unresolved stage of a product challenge.
Product Router
identifies the earliest unresolved stage
Explore
Understand ambiguous problems and identify opportunities worth pursuing.
Review
Challenge existing ideas, evaluate assumptions, and expose weaknesses.
Plan
Translate grounded direction into priorities, strategies, and actionable decisions.
Build
Apply structured reasoning to experiences, workflows, and implementation decisions.
Rather than forcing every request through a rigid process, the system adapts its reasoning depth to the situation.
THE ARCHITECTURE
Building relationships, not collecting frameworks.
The central architectural challenge was teaching the system how knowledge connects.
I structured a shared knowledge graph linking problem classes, mechanisms, outcomes, strategies, risks, metrics, validation methods, and examples.
Instead of mapping a symptom directly to a recommended tactic, Preflight uses these relationships to investigate why a problem exists and which interventions the evidence supports.
THE KNOWLEDGE MODEL
Symptom → Tactic
Symptom → why → what the evidence supports
Problem classes
What kind of problem is this?
Mechanisms
Why does it exist?
Outcomes
What needs to change?
Strategies
Which interventions does the evidence support?
CONNECTED THROUGHOUT
Risks
Metrics
Validation methods
Examples
627
REUSABLE KNOWLEDGE OBJECTS
924
CONNECTED RELATIONSHIPS
The value isn't the volume of knowledge. It's the relationships between it.
Diagram: simplified view of how Preflight's knowledge types relate.
DESIGN DECISIONS
Designing for judgment, not just answers.
Three principles shaped how Preflight reasons.
01
Investigate mechanisms, not just symptoms.
Different causes can produce identical symptoms. Preflight explores alternative explanations before recommending an intervention.
02
Match confidence to evidence.
The system distinguishes facts, interpretations, assumptions, and predictions, making uncertainty explicit rather than disguising it with confident language.
03
Make reasoning proportional.
Not every problem needs an exhaustive investigation. Preflight retrieves relevant knowledge and adjusts its depth to the complexity of the decision.
APPLICATION
Sometimes the interface is only where the problem appears.
Consider a team responding to complaints that users don't know when an approval will finish.
The obvious solution is a status indicator. But the real issue might be unclear ownership, missing service expectations, an absent escalation path, or inconsistent governance.
Preflight helps teams investigate those possibilities before committing to an interface change.
ILLUSTRATIVE REASONING EXAMPLE
OBSERVED SYMPTOM
"Users don't know when approval will finish."
ASSUMED SOLUTION
ALTERNATIVE CAUSES TO INVESTIGATE
Unclear ownership
No defined SLA
Missing escalation
Ambiguous workflow states
REASONED DIRECTION
Determine the underlying cause before choosing the intervention.
The result isn't necessarily a different interface. It's a better-informed decision about whether the interface needs to change at all.
VALIDATION
Testing whether the system could reason, not just respond.
I iteratively tested Preflight against product challenges spanning AI opportunities, workflow automation, discovery, and planning.
The evaluations examined mode selection, knowledge retrieval, diagnostic reasoning, evidence handling, and the usefulness of recommendations.
Each round informed refinements to routing, knowledge relationships, and response behavior, establishing a more consistent baseline for future development.
FROM TESTING TO REFINEMENT
TEST SCENARIO
A product planning request requiring strategic direction before moving into implementation.
BEHAVIOR OBSERVED
Preflight introduced Build-oriented guidance before the strategic planning decisions had been resolved.
REFINEMENT MADE
Clarified routing rules and mode boundaries to keep strategic reasoning in Plan until implementation was appropriate.
OUTCOME
From concept to functioning AI capability.
Preflight evolved into a functioning internal system built on Claude, combining four specialized reasoning modes with a shared knowledge architecture.
The work established a reusable foundation for investigating product problems, evaluating solutions, and supporting decisions across product, design, engineering, and strategy.
Its implementation and scenario testing demonstrated the feasibility of the approach, while broader adoption and measurable business impact remained future opportunities.
REFLECTION
Better AI doesn't just produce better answers. It helps us ask better questions.
Preflight reinforced something central to how I approach innovation: the hardest part of solving a problem is often understanding what the problem actually is.
The most meaningful design work wasn't creating hundreds of knowledge objects. It was determining how they needed to connect, when they should be applied, and how to preserve human judgment throughout the process.
NEXT UP
Reimagining a fragmented news experience across 70+ local markets, balancing audience needs, business objectives, and the realities of an established publishing ecosystem.