↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Spec-driven development/

AI Unified Process

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 6 minutes read (~1031 words)
agent-curated 550 fully-ai-generated 653 llm=glm-5.3-flash 561 methodology 3 open-source 176 research-note 520 rup 1 spec-driven-development 19 use-cases 1
Table of Contents

AI Unified Process (AIUP) is Simon Martinelli’s requirements-first, Apache-2.0 methodology that keeps use cases as the living source of truth, generates code and tests from them with coding agents, and audits the result back against the spec, distributed as agent plugins and skills for Claude Code, Codex, Cursor, Copilot, Gemini CLI, and OpenCode.

AIUP is this category’s clearest inversion: where every other column treats the spec as input to code, it declares the code the current implementation of the spec, and it arrives with the category’s first published textbook (the steward’s own Apress book on the method) and a Rational Unified Process lineage.

What it is
#

An open-source plugin set (aiup-core plus four stack plugins) installed per agent from its marketplace, through the Tessl registry, or as manual skill files. aiup-core owns the stack-independent path: /requirements builds the catalog from docs/vision.md, then /entity-model, /use-case-diagram, /use-case-spec (which asks up to five questions whose answers change the spec before writing), /business-process, /test-case, /spec-review, and /reverse-engineer for existing codebases. A stack plugin (Vaadin/jOOQ, Angular/JPA, Blazor/.NET, or NestJS/Next.js) continues from those artifacts into migrations, implementation, and tests. Artifacts carry stable identifiers (FR-*, UC-*, BR-*, TC-*) referenced downstream once, with code markers such as // UC-005 BR-003: ... and test annotations tying every business rule to its implementation. The four stack plugins are uneven: only Vaadin/jOOQ and Angular/JPA ship the read-only uc-coverage sub-agent and the /coverage-check skill that judge implementation and tests together.

Status
#

Active and authored: 146 stars, 26 forks, and 6 open issues since creation on 2026-02-28, pushed 2026-10-17, Apache-2.0, with per-plugin version tags (aiup-vaadin-jooq-v2.21.0 the newest) as of 2026-10-17 (GitHub API).

Star History Chart

The adoption case rests on named field use rather than stars: reference customers (WBS GRUPPE, Zooey GmbH, centeractive), one community-contributed stack plugin, an independent endorsement in the Ask HN spec-driven thread (2026-09-09), and a public third-party field report (a 13-use-case case-management build whose coverage audit caught three gaps the agent then closed). It has no Hacker News thread of its own, and the name collides with an unrelated aiup.dev story, so searching for coverage takes care. Martinelli self-places it at “spec-anchored, not spec-first”, the middle level of Birgitta Böckeler’s three-level framing of spec-driven development.

Strengths
#

  • The code-serves-the-spec inversion is carried by machinery, not prose: spec_lint.py fails CI on broken identifiers, requirement status is derived from the use cases with a drift warning, and /coverage-check maps every use-case unit onto the code and tests that realize it.
  • Traceability is bidirectional by design: references are written downstream once and the reverse direction is computed, so “why does this exist” is one command (spec_lint --trace --only FR-014).
  • Brownfield is a first-class workflow: /reverse-engineer recovers entity and use-case baselines from an existing codebase.
  • Steward credibility: a Java Champion training teams on exactly this method, with named reference teams, a tutorial, videos, and a book behind it.

Cautions
#

  • 146 stars is among the smallest adoption figures in this category, and the visible adoption sits inside the German-speaking Java and Vaadin consulting orbit the stack plugins encode.
  • The full artifact chain (vision, catalog, entity model, diagram, use cases, processes, test cases) is heavyweight, nothing sizes it down for small changes, and the field report is blunt that in production work the review time must actually be spent.
  • The coverage audit exists in only two of the four stack plugins; other stacks run the same check by hand from the agent definition.
  • Solo steward with paid training and a private-beta Studio around the method, and Apache-2.0 with NOTICE and trademark terms on derived works, unlike this category’s MIT norm.

Pricing
#

The plugins and the Navigator IDE plugin are free and open source under Apache-2.0. Training is the paid surface: an in-house workshop at CHF 4,400 per day for up to 10 people, and online sessions from $290 (as of 2026-10-17). AI Unified Studio, the web workspace for non-developer stakeholders, is private beta by invitation with no published price, and the Apress book is sold separately; the Leanpub use-case guide is free.

Price history
#

Date Plan Change Source
2026-10-17 All Baseline at creation: plugins and Navigator free (Apache-2.0), in-house workshop CHF 4,400 per day up to 10 people, online sessions from $290, Studio unpriced in private beta. unifiedprocess.ai

Compared to
#

  • BMad Method: the other whole-method column; BMad right-sizes the ceremony and stays code-first, AIUP runs a fixed requirements-to-construction chain with the spec on top.
  • GitHub Spec Kit: scaffolds conventions per change and leaves the spec behind; AIUP keeps the spec anchored for the system’s life and reconciles code to it.
  • OpenSpec: versions the spec as change deltas; AIUP versions behavior as use cases and treats code as the derivative.

Bottom line
#

Recommended for analyst-plus-developer teams on the supported stacks, especially modernizing business applications where the requirements should outlive the framework. Not for solo or small-change work, and not for stacks without a plugin, where the method ends at documented specifications. The disagreeable claim I will defend: spec-kit taught the industry to write specs before code, but AIUP’s direction, code as the current implementation of a living spec, is where this category’s survivors are heading, and the book-plus-audit combination is the first serious attempt to make that teachable.

Changes
#

  • 2026-10-17 - Created from the Engineering4AI awesome-list entrant scan, profiling AI-Unified-Process/marketplace as the category’s requirements-centric, RUP-lineage column.

See also
#

References
#