TL;DR. Four meta-skill frameworks that bolt onto Claude Code and force an opinionated way of working entered the same spot in succession. Superpowers (153k★, 2025-12) · GSD (61.9k★, 2025-12) · ECC (181.6k★, 2026-01) · gstack (95.9k★, 2026-03). Each constrains a different axis — Superpowers locks process gates, GSD enforces context separation, gstack imposes role personas, ECC opinionates the harness baseline. Which one (if any) you install, and how you mix them, becomes the central workflow decision of the next quarter.
What is an Agent Framework
Definition — a meta-skill bundle that bolts onto Claude Code (or an equivalent coding agent) and forces an opinionated way of working as the default. Not a simple skill pack, but a methodology.
The difference from a solo-author signature skill pack is the crux:
| Dimension | skill pack | agent framework |
|---|---|---|
| Unit | one kind of task | the whole order, discipline, and roles of work |
| Invocation | slash command or description match | bootstrap hook — always on |
| User decision | decide which skill to call each time | install once, auto-applied every session |
| Analogy | toolbox | work OS |
A framework is a methodology, not a plugin — it rewrites Claude Code's default behavior itself.
What showed up all at once
In roughly four months, from December 2025 to March 2026, all four frameworks hit GitHub's trending top tier and pulled in tens to hundreds of thousands of stars in short order.
- 2025-12 — obra/superpowers. Jesse Vincent's 14 meta-skills + a SessionStart bootstrap. It binds brainstorm → plan → SDD → review · TDD · debugging · verification into a self-invoking pipeline. Currently 153k★.
- 2025-12 — gsd-build/get-shit-done (GSD). TÂCHES's spec-driven development system. A five-phase cycle — discuss → plan → execute → verify → ship — with a fresh 200k-context subagent per phase. Supports 15 runtimes (Claude Code, Codex, Gemini CLI, Cursor, …). 61.9k★, 2,500+ commits, 136 contributors.
- 2026-01-22 — affaan-m/everything-claude-code (ECC). A harness-performance optimization system by Anthropic hackathon winner Affaan Mustafa. With 60+ agents + 220+ skills + 75+ commands, it opinionates the harness baseline (tokens, memory, security, parallelism). The only one in the category with a commercial-platform plane (hosted SaaS, GitHub App, separate npm package, free/Pro/Enterprise marketplace). 181.6k★, 28k forks, 1,900+ merged PRs.
- 2026-03-11 — garrytan/gstack. Y Combinator CEO Garry Tan opened his own Claude Code setup as-is. 23+ slash commands each map to a role — CEO, Designer, Eng Manager, QA, Release Engineer, Doc Engineer, Security Officer. 10k★ in 48 hours, 50k★ in 16 days, currently 95.9k★.
What the four share is that they rewrite Claude Code's default behavior in an opinionated way. What differs is which axis they rewrite.
Where the four frameworks stand
| Framework | What it constrains | Author | Metaphor | Stars |
|---|---|---|---|---|
| Superpowers | process gates (brainstorm, plan, TDD, review) | obra (Jesse Vincent) | work OS | 153k+ |
| GSD | context separation (fresh subagent context per phase) | TÂCHES | context compiler | 61.9k+ |
| gstack | role division (CEO, Designer, QA, Release personas) | Garry Tan (YC) | virtual company | 95.9k+ |
| ECC | harness baseline (tokens, memory, security, parallelism) | Affaan Mustafa (zenith.chat) | harness OS + commercial platform | 181.6k+ |
Same category, different constraint
All four are opinionated, but the axis of the opinion differs.
- Superpowers has an opinion about how to work. Brainstorming before every code change, TDD for every new feature, systematic debugging for every debug — it makes process discipline the first-class citizen. The order and the gates of work are the substance.
- GSD has an opinion about how to cut up context. It spins up a fresh 200k-context subagent for each phase of discuss/plan/execute/verify/ship — it solves context rot with physical separation. The boundaries of work are the substance.
- gstack has an opinion about who works. A single coder cycles through 23 roles. Review the product in CEO mode, inspect the UI in Designer mode, run an OWASP+STRIDE audit in Security Officer mode. The actor of the work is the substance.
- ECC has an opinion about how to run the harness itself. Token defaults (model: sonnet · MAX_THINKING_TOKENS=10000 · autocompact=50), memory persistence (SessionStart/Stop hooks bridging across sessions), instinct-based continuous learning, AgentShield security scanning, parallel worktree operation — it makes Claude Code's execution environment itself the first-class citizen. The baseline of work is the substance.
The four axes are orthogonal. You can install them together, and they can conflict — where they collide is covered in each framework's standalone piece.
Where it will break
Given that all four arrived at once, the operational themes for next quarter narrow to four:
- Bootstrap collision. Superpowers, GSD, gstack, and ECC all want to grab the SessionStart hook. Install two or more and which meta-skill lands in context first dictates the behavior.
- Name collision. All four have slashes and skills in the TDD, debugging, review space. That's where description matching splits.
- Role vs gate mismatch. gstack's role personas are orthogonal to Superpowers' process gates — install both and a conflict like a CEO role bypassing the brainstorming gate becomes possible.
- Catalog overwhelm vs process gate. ECC's 220+ description-rich skill catalog can bypass Superpowers' process gates — if an ECC vertical skill matches a work utterance first, the task starts before the brainstorming HARD-GATE can fire.
Related pieces
Each framework covered on its own — what it is, why install it, how to use it, and where it collides by design when mixed with the others — deduced from hooks and definitions, not from running them together.
- Superpowers — at v5.1.0, the de facto standardization of 14 meta-skills + a SessionStart bootstrap
- GSD — at v1.41.0, how the five phases (discuss, plan, execute, verify, ship) spin up a fresh 200k context each phase to solve context rot at the environment level
- gstack — at v1.34.1, how 23 slashes + 8 power tools map one-to-one onto CEO, Designer, Eng Manager, QA, Release Engineer, and Security Officer roles, turning a solo coder into a virtual company
- ECC — at v1.10 + the 2026 Q2 fix-wave, the commercial platform where 60+ agents + 220+ skills + 75+ commands redesign Claude Code's baseline along five axes: tokens, memory, instinct, AgentShield, and parallelism