When code costs almost nothing, the plan becomes the product. That's the core thesis behind Digital Innovation Agents (DIA), an open-source workflow toolkit that drops a V-Model development process directly into your AI coding assistant of choice. The project dropped on GitHub this week with thirteen skills covering Business Analysis, Requirements Engineering, Architecture, Coding, Testing, and Security Audit—all wired together with quality gates between every phase transition.

What You're Actually Getting

DIA isn't another prompt library or collection of loose agentic workflows. It ships as a structured plugin system that enforces sequential handoffs: each phase skill produces an artifact (ADRs, feature inventories, test suites) and won't let you proceed to the next stage until its quality gates pass. The approach mirrors traditional V-Model development but strips out the ceremony while keeping the traceability. Every requirement traces back to a real user problem; every code change traces forward through tests and security review. The skill set splits into three groups: six V-Model phase skills that own their respective stages, four foundation skills handling project conventions and consistency checks, and an orientation bootstrap that loads automatically on session start. Phase skills include /business-analysis for structured discovery, /requirements-engineering for epic and feature decomposition, /architecture for ADR generation in MADR format, /coding with PLAN-NN persistence and coverage gates, /testing enforcing AAA pattern and FIRST principles, and /security-audit covering OWASP Top 10 and LLM Top 10.

The Innovation Methodology Angle

Here's where DIA gets interesting beyond just workflow automation. The BA and RE agents ship a catalog of thirty-two field research methods—qualitative interviews, extreme user studies, fly-on-the-wall observation, TRIZ, pre-mortem analysis, value proposition quantification, and more. During discovery sessions, when your answers go thin, the agent stops and proposes the matching method with a one-page card: what it produces, when to use it, how to run it, team size and time estimates, common failure modes, and what deliverables to bring back. The catch? The actual research work stays human-to-human. Interviews happen between real people. Observations occur in real contexts. Prototypes land on real hands. The agent's job is spotting when you're guessing instead of knowing—and pointing you toward the right technique to fill the gap. This is less "AI does your job" and more "AI keeps you honest about whether you've actually talked to users."

Cross-Platform Support

DIA targets Claude Code as its recommended platform, with official plugin marketplace support via /plugin commands. But it also ships installers for Cursor, GitHub Copilot (CLI and VS Code), Codex, OpenCode, and Gemini CLI—each platform gets the same thirteen skills with identical templates and quality gates. For Claude Code, installation runs through a single curl pipeline; for Copilot, agents copy into .github/ subdirectories so helper scripts resolve relative to project root. The consistency-check skill runs in three modes: syntactic (dead links and orphan features), semantic (requirement-to-code alignment), or full audit. It's mandatory at every phase boundary by design—no skipping ahead because you feel confident about the architecture when you've skipped user research entirely.

Design Principles Worth Noting

DIA enforces tech-agnostic success criteria in requirements: no OAuth, REST, PostgreSQL, or React vocabulary allowed in the contract between user and team. Technical details live in separate NFR sections and ADRs that follow during architecture. This separation means stakeholders read what the system does; engineers decide how it does it. Living documentation is another core principle. ADRs, features, architecture docs, and backlog update continuously during implementation. By release time, documentation reflects what was actually built—not a snapshot of original intentions. The _devprocess/context/BACKLOG.md file serves as single source of truth for project state.

Key Takeaways

  • Thirteen skills with quality gates enforce V-Model discipline across six development phases
  • Thirty-two innovation methods keep teams honest about user research rigor
  • Tech-agnostic requirements separate "what" from "how" throughout the lifecycle
  • Cross-platform support spans Claude Code, Cursor, Copilot, Codex, OpenCode, and Gemini CLI

The Bottom Line

DIA is MIT-licensed by developer Sebastian Hanke with active v3 development. For teams drowning in sprint velocity metrics while shipping features nobody asked for, this framework offers a structured alternative that treats evidence of user value as non-negotiable—not a nice-to-have after the code ships.