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

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).