Claude Code skill packs: superpowers vs GStack vs get-shit-done vs Anthropic Skills
A skill is a folder of instructions (a SKILL.md file, plus any scripts it needs) that a coding agent loads only when the task matches, instead of carrying every instruction all the time. A skill pack is a curated bundle of them: someone else's working habits for your agent, installed as files. Nothing to run, nothing to migrate off, which is why How to pick a harness says to start at this tier.
The cost is behavioral, not financial. A pack rewrites how your agent plans, tests, and commits, so the wrong pack means weeks of fighting a workflow you didn't choose, and stacking two packs that disagree is worse than running neither. One published test of popular skills found many made output worse. Install one, watch what changes, prune what never fires.
If you're deciding between a skill and the other extension points, the split is: files like CLAUDE.md hold what the agent must always know, skills hold procedures loaded on demand, subagents are separate workers with their own context, and MCP servers add external tools (Anthropic's own explainer covers this; so does our context files page). The four packs below are different bets on what belongs in the skill folders.
| superpowers | GStack | get-shit-done | Anthropic Skills | |
|---|---|---|---|---|
| ⭐ Stars | 289k | 134k | 9.7k | 177k |
| The bet | Process: test-driven development, systematic debugging, verification before claiming done | Roles: 23 slash-command modes (CEO review, eng review, design, QA, ship) that structure one assistant as a virtual team | Plans: goal-backward plans on disk, executed in waves over fresh context windows | The format: the official reference skills, plus document production (docx, pdf, pptx, xlsx) |
| Runs on | Claude Code plus 13 other harnesses (Codex, Cursor, OpenCode, Gemini CLI, more) | Claude Code | Claude Code, OpenCode, Gemini CLI | Claude Code, Claude.ai, the API |
| How it enforces itself | A startup hook re-injects its rules when a session begins and after compaction (when a long session gets compressed and standing instructions usually fall out) | Checkpoint mode auto-commits work as it goes; /freeze and /careful guardrails; its browser tool only allows pre-approved commands | The plan is a file on disk, so execution survives a dead session and picks back up | It doesn't: pure content, the host platform decides |
| Memory between sessions | Basic: an optional companion plugin adds recall | Strong: per-repo learnings plus an optional GBrain knowledge backend | The plan files are the memory | None |
| License | MIT | MIT | MIT | ⚠️ Per-skill Anthropic terms ("all rights reserved") |
Stars as captured for the main list; rating definitions live in the guide to rankings. Enforcement and memory rows come from this list's deep-dive research and each repo's own documentation.
Pick by situation
- You want the agent to work like a disciplined engineer → superpowers. The most-starred pack of the category. Its skills are process (write the test first, debug systematically, verify before claiming done), and its startup hook re-injects that discipline exactly when packs usually lose it: after compaction.
- You want a product team, not just an engineer → GStack. Garry Tan's daily-driver stack turns one assistant into reviewers, QA, and a ship pipeline, with the strongest memory story of the four: what a project teaches the agent feeds its later sessions.
- Your projects outlive single sessions → get-shit-done. Its whole design is plans as files: work is planned backward from the goal, executed in fresh context windows wave by wave, and a crash or a cleared session resumes from the plan instead of starting over.
- You're learning the format, or you need documents produced → Anthropic Skills. The first-party reference implementations everything else builds on, plus the docx/pdf/pptx/xlsx workhorses. Read the licensing before you redistribute anything: these are not standard open source; each skill ships an all-rights-reserved Anthropic-terms license file.
They compose, up to a point
Skills are folders, so nothing stops you from installing two packs, and a combining guide exists for exactly the trio above. The failure mode is trigger overlap: two packs claiming the same moment ("before any bug fix, use MY debugging skill") leave the model choosing between rulebooks mid-task. Start with the pack whose bet matches your gap, add single skills from the others, and prune anything that never fires.
Beyond these four: addyosmani/agent-skills (84.9k stars) optimizes for portability, with 24 senior-dev workflow skills installable across 70+ agents; ECC is the breakout mega-pack (68 specialized subagents and 284 skills at last count); wshobson/agents is the cross-harness marketplace of drop-in agent definitions; and awesome-claude-code is the most-followed catalog of the whole genre, the place to discover packs this page doesn't name.
Part of best-of-Agent-Harnesses. New to this decision? Start with How to pick a harness. Spot an error or a stale claim? Open an issue.
Agents can query this page's data directly: claude mcp add agent-harnesses -- uvx agent-harnesses-mcp (MCP server).