23 AI roles, a 7-step standard pipeline, and a claimed 810x efficiency boost — these are the core selling points of gstack, an open-source toolkit. It repackages Claude Code from a general-purpose assistant into a virtual engineering team.
What this is
gstack is an open-source toolkit that decomposes a single AI coding agent into 23 specialized roles: CEO (re-examines product requirements), Engineering Manager (locks in architecture), Designer (blocks low-quality UI), QA (operates a real browser to open pages and verify), Security Officer (runs OWASP checklists and STRIDE threat-modeling audits), Release Engineer (submits merge requests), and more. Each role is invoked via slash commands (e.g., /qa, /browse), and the work is forced through a seven-step progression — Think → Plan → Build → Review → Test → Ship → Reflect — with each step's output automatically handed off to the next.
It is not locked to Claude Code. Via the --host parameter, it also supports OpenAI Codex, Cursor, GitHub Copilot, and other coding environments. On the design side, /design-shotgun lets GPT Image generate multiple UI sketches at once for selection; on the security side, there are prompt-injection detection and destructive-command blocking mechanisms.
Industry view
The strongest argument from author Garry Tan is the 810x figure — he claims that with gstack, his "logical code change velocity" is 810 times that of 2013 (11,417 lines per day vs. 14). The number is sticky, but it has two soft spots worth flagging. First, "logical code change" is a metric he defined himself, already excluding redundant portions of AI-generated code — essentially setting his own rules and reporting his own results. Second, the baseline stops at 2013 — a time before GitHub Copilot, GPT-4, or cloud-based AI IDEs existed. Comparing today's output against the daily averages of a decade-plus ago is structurally biased toward the new tool.
Supporters will argue that gstack's real value isn't in that number, but in upgrading AI coding from a "single-soldier tool" to a "process-driven team" that makes the methodology reproducible. Critics will counter that piling up tools doesn't equal engineering capability — stringing 23 roles together doesn't inherently produce good software; if requirement decomposition and architectural judgment are off-base, AI just ships bad code faster.
Impact on regular people
For enterprise IT: If your company has outsourced dev teams or is evaluating AI coding procurement, "process-driven collaboration" is becoming a new selling point, and the methodology itself may start appearing on procurement lists.
For individual careers: Non-technical roles won't need it for now, but the "AI role-based division of labor" idea will gradually seep into marketing, customer service, content production, and other scenarios — expect to see this phrase repeatedly over the next few years.
For consumer markets: No direct short-term impact. If tools like gstack mature, the cost for independent developers and small teams to ship small SaaS products will keep falling, and more niche micro-tools will get built.