IntentSpec: Explicit Interpretation and Independent Verification for AI-Accelerated Requirements Engineering
Abstract
AI coding agents can transform software requirements into implementations rapidly, but faster implementation does not guarantee correct interpretation. Conventional natural-language requirements often leave authority, ambiguity, traceability, and verification evidence implicit, while highly formal approaches may impose substantial authoring overhead.This paper presents IntentSpec, an experimental Markdown-based requirements representation and companion agent workflow designed for joint use by human engineers and AI coding agents. IntentSpec separates normative obligations from informative context, provides stable requirement and acceptance-criterion identifiers, supports typed traceability and structural approval gates, and requires unresolved questions to remain visible. A proposed interpretation protocol asks agents to externalize triggers, obligations, constraints, assumptions, and ambiguities before implementation, while verification is treated as a separate activity requiring independent evidence rather than agent self-attestation.A working Python prototype implements structural parsing, validation, cross-reference checking, JSON export, conformance tests, and CI validation. We outline a controlled evaluation against plain Markdown, structured natural-language requirements, and an existing specification-driven agent workflow, measuring interpretation accuracy, ambiguity detection, implementation compliance, traceability, and human authoring effort.
