AI-first product
Turning discovery into a specification engineering can trace
SpecFlow explores how AI can shorten the distance between discovery notes and product decisions ready for execution while preserving context, limits, and evidence.
Context
The problem was not writing one more document.
Discovery conversations contain symptoms, hypotheses, constraints, and phrases that give the problem meaning. When this material is summarized without traceability, a specification may look clear while losing the context that would help a team decide.
- Preserve conversation evidence
- Separate hypothesis from requirement
- Make the MVP boundary explicit
Product question
How can synthesis speed up without turning ambiguity into a requirement?
The direction was to treat generation as an assisted structuring step. The user receives the problem, hypothesis, boundaries, traceable stories, backlog, and risks in one artifact that can be reviewed before execution.
- Generate a useful first structure
- Allow human review before use
- Keep literal quotes recoverable
Architecture decision
The AI output needs to be a contract, not a block of text.
The product combines generative providers, a Zod-validated JSON contract, integration fallback, and staged loading that makes the process legible. This gives the interface a unit of work closer to an engineering document than a loose answer.
- Strict specification schema
- Fallback between Gemini and OpenRouter
- Markdown reader and export
Evaluation
Grounding is the first test of usefulness.
The harness checks whether generated quotes actually exist in the input, measures scope completeness, and penalizes ambiguity. The first format-validation run reached 91% grounding and reduced first-review time by 32%.
- 91% grounding in evaluated quotes
- 32% less time in first review
- 86% scope completeness
Conceptual architecture
From conversation to an executable contract, with traceability at every step.
The flow turns discovery signals into a validated specification, switches providers, and checks grounding before delivery.
Discovery
Raw notes, transcripts, and business context.
Structure
Problem, hypothesis, scope, and traceable stories.
Providers
Switchable models with operational fallback.
Contract
Validated JSON output before it reaches the interface.
Evals
Grounding, completeness, and ambiguity signals.
Featured learning