Six months after Anthropic shipped a plugin system for Claude Code, the most-installed extensions on the platform don’t have anything to do with new capabilities. They’re opinionated rulebooks — and the three which have damaged into mass adoption this spring share a thesis that has since unfold to 14 completely different AI coding brokers: the bottleneck in AI-assisted software program growth isn’t mannequin intelligence. It’s self-discipline.
GSD, the youngest of the three frameworks, marked its 57th launch in as many weeks on Could 3 and has since shipped additional updates, with 61,600-plus GitHub stars and 138 lively contributors. Alongside Superpowers’ 192,000-plus stars and gstack’s near-100,000 stars, the figures define a class that hardly existed on the finish of 2025 and now shapes how a whole bunch of 1000’s of builders work together with AI coding brokers every day.
Superpowers: Self-discipline as Iron Regulation
Superpowers, constructed by Jesse Vincent — the Perl 5 launch supervisor, Keyboardio co-founder, and creator of the Request Tracker ticketing system — launched on October 9, 2025, the identical day Anthropic opened Claude Code to third-party plugins. Anthropic accepted it into the official plugin market on January 15, 2026. As of Could 2026 it carries 192,000-plus GitHub stars, 17,100 forks, and a v5.1.0 stable release.
The framework ships as a folder of markdown information — 14 ability information, every encoding a section of the event cycle: brainstorm, plan, implement, evaluate, ship. The important thing design alternative isn’t what these phases comprise however how they’re enforced. Each ability file opens with what open-source maintainer Marc Nuri, a Senior Principal Software program Engineer at Pink Hat who reviewed the framework intimately, described as a “capitalized, non-negotiable rule, an Iron Law, followed by a table of red flags: the rationalizations the agent is most likely to use to skip the rule.”
That construction targets a selected failure mode: the tendency of enormous language fashions to purpose their means out of constraints mid-session. The goal, Nuri wrote, isn’t educating the agent — as a result of it already is aware of the principles — however stopping it from speaking itself out of following them. Simon Willison, the creator of Datasette and co-creator of Django, referred to as Vincent “one of the inventive customers of coding brokers” he is aware of.
The sensible end result: when a developer asks Claude Code to construct one thing with Superpowers put in, the agent doesn’t write code. It opens a Socratic design dialog, forces approval of a spec, then proceeds by necessary test-driven growth — crimson, inexperienced, refactor, no exceptions. A tough gate blocks implementation till the planning section is full.
gstack: Self-discipline as Position Separation
Y Combinator President and CEO Garry Tan open-sourced gstack on March 12, 2026. It hit 50,000 GitHub stars in 16 days — TechCrunch covered the launch — and has since climbed towards 100,000 stars with 284 commits and 49 contributors.
The place Superpowers constrains what the agent does inside a job, gstack constrains which perspective the agent occupies earlier than a job begins. The framework splits a single Claude Code session into 23-plus named roles — CEO, Designer, Engineering Supervisor, Launch Supervisor, QA, Doc Engineer, Chief Safety Officer — every carried out as a slash command with its personal priorities and constraints. The premise, drawn immediately from Tan’s README, is that “a single builder with the best tooling can transfer quicker than a standard crew” — however provided that the agent is compelled to modify modes between product, engineering, and high quality work moderately than mix them.
Tan claims to have averaged roughly 11,400 logical traces of code per day over a current 60-day interval, describing that as roughly 810 occasions his 2013 tempo, with the methodology and a copy script printed within the repository itself for verification. Unbiased reviewers have been skeptical however not dismissive: line counts are a weak proxy for code high quality, and gstack’s most credible worth proposition is structural — the role-switching workflow — not its creator’s private output figures.
The reception was polarized virtually instantly. YouTuber Mo Bitar produced a video titled “AI is making CEOs delusional,” arguing that gstack is basically “a bunch of prompts in a textual content file.” On Product Hunt, startup founder Sherveen Mashayekhi wrote that if Tan weren’t the CEO of Y Combinator, the challenge wouldn’t have been featured there.
These critiques should not fully unsuitable. gstack is markdown information. Any skilled Claude Code developer has seemingly assembled some model of this workflow privately. However that misses the purpose that a number of reviewers converged on: the worth of gstack isn’t proprietary expertise. It’s a battle-tested, opinionated workflow encoded as soon as moderately than reinvented per challenge — and co-developed with 49 contributors utilizing the identical Claude Opus mannequin the framework is designed to information.
GSD: Self-discipline as Context Hygiene
GSD — Get Shit Performed — takes the prognosis in a 3rd path. Created in December 2025 by developer Lex Christopherson, who publishes below the names TÂCHES and “glittercowboy,” it targets what it calls context rot: the standard drop that happens as an AI coding session fills its context window.
The mechanism is architectural. Moderately than constraining what an agent does or what function it occupies, GSD breaks work into atomic plans, then executes every plan in a recent sub-agent session with a clear 200,000-token context window. The primary session — used just for orchestration — is stored at 30 to 40 % of its window all through the challenge. The objective is that job 50 maintains the identical high quality as job 1.
GSD has shipped 60-plus releases since December 2025. The v1.40.0 replace on Could 3 added a Minimal Set up Profile that cuts system immediate overhead from roughly 12,000 tokens to 700 — a 94 % discount that makes GSD viable for native fashions and metered API plans. The v1.42.x sequence, launched through the week of Could 11, added per-phase mannequin choice, dynamic routing with failure-tier escalation, and a bundle legitimacy gate towards what the discharge notes name “slopsquatting” — malicious packages designed to impersonate legit dependencies.
The framework now helps 14 AI coding brokers — Claude Code, Cursor, Codex, Copilot, Gemini CLI, Windsurf, Cline, Increase, and 6 others — and ships an installer that auto-detects and configures the best file structure for every runtime.
Three Diagnoses, One Convergence
What makes the second notable isn’t any single framework however their convergence on the identical underlying argument. Not one of the three provides mannequin functionality. Superpowers is essentially a system of prohibitions. GSD describes itself as a “context engineering system.” Garry Tan, in posts surrounding gstack’s launch, framed his instruments explicitly as compelled cognitive switching, not intelligence amplifiers.
The implicit declare throughout all three is that present frontier coding fashions are routinely succesful sufficient to ship manufacturing code — and routinely talked out of doing it nicely by the best way they’re prompted. Every framework codifies a unique reply to which intervention issues most: rule self-discipline (Superpowers), function separation (gstack), or context hygiene (GSD). The neighborhood has begun treating them as complementary layers moderately than rivals — gstack for strategic choices, GSD for context stability, Superpowers for execution — with unified set up guides and mixture workflows rising throughout developer blogs and documentation.
Anthropic has, at minimal, implicitly endorsed the path by that includes Superpowers as a verified market plugin and by consolidating its personal slash instructions and abilities programs right into a single unified abilities format — a change that makes community-built ability frameworks the canonical extension mechanism for Claude Code.
Adoption Caveats and What Comes Subsequent
The adoption numbers carry caveats. GitHub stars and set up counts measure consideration, not outcomes. GSD’s declare of use by engineers at Amazon, Google, Shopify, and Webflow seems in its personal README and has not been independently confirmed by these firms. Tan’s line-count figures have drawn constant criticism from builders who word that uncooked output quantity is a poor proxy for shipped worth.
The trajectory can also be forking in ways in which recommend the present plugin structure is an intermediate state. GSD’s v2 TypeScript rewrite moved from markdown prompts to a TypeScript software particularly as a result of, as its maintainers famous, injecting directions by slash instructions left no precise management over context home windows, classes, or execution move. The v1.42.x sequence provides per-phase mannequin choice and dynamic routing — capabilities that sit nearer to infrastructure than to prompting.
That shift — from discipline-as-prompting to discipline-as-infrastructure — marks the seemingly subsequent section of this class. If it succeeds, the methodology will now not depend upon an agent selecting to comply with guidelines. It will likely be the runtime.
For now, the less complicated studying is the one Superpowers retains pushing. The brokers already know find out how to write code. They should be instructed, in writing, to not begin till they’ve deliberate.