External skill and agent frameworks
What exists, what it ships, and whether it fits a lean solo setup that runs Claude Code, GitHub Copilot and OpenCode from one source. Reviewed 2026-10-12; star counts and plugin numbers are approximate.
| Framework | Ships | Clients | Fit | We use |
|---|---|---|---|---|
| angular/skills (official) | angular-developer: one router SKILL.md plus ~30 references; angular-new-app | all skills clients | Adopt. Progressive disclosure done right; overlaps ng-implementation. Generic, no project conventions. APM path: angular/skills/angular-developer#<sha>. | planned (#94) |
| dotnet/skills (Microsoft) | 17 plugins under plugins/; dotnet = 4 skills + C# LSP, dotnet-test = 23 skills + 10 agents | Copilot CLI, Claude Code, VS Code, Cursor, Codex | Adopt plugins/dotnet only. dotnet-test floods the skill list. Pin with #<sha>. | decision open (#94) |
| anthropics/claude-plugins-official | ~300 plugins: commit-commands, pr-review-toolkit, code-review, feature-dev, skill-creator, security-guidance, LSP plugins | Claude Code only | Selective: skill-creator, pr-review-toolkit (compare with code-review on one PR). /plugin shows the context cost of each. Nothing for Copilot or OpenCode. | user-local |
| anthropics/skills | docx/pdf/pptx/xlsx, frontend-design, mcp-builder, skill-creator; the spec | Claude Code, claude.ai, API | Reference for skill authoring; no overlap with the workflow skills. | via /plugin |
| obra/superpowers | 15 skills: TDD, brainstorming, writing-plans, subagent-driven-development, systematic-debugging; SessionStart bootstrap; v7.0.0 (2026-10-09) | Claude, Codex, Cursor, OpenCode, Copilot CLI, Gemini, … | Borrow, not adopt (#96 closed without the A/B run, see Evaluations). Four mechanics copied into the agents (#113); no issue or release-branch model, no stack skills, mandatory bootstrap per session. | borrowed |
| github/awesome-copilot | ~100 agents, skills, instructions, plugins (Gem Planner/Implementer/Reviewer, .NET agents) | Copilot CLI/VS Code; installable with npx skills | Cherry-pick single agents by path if ever needed; agents are verbose. | no |
| wshobson/agents | 94 plugins, 202 agents, 184 skills; one Claude source + adapters for Codex/Cursor/OpenCode/Copilot | Claude Code primary | Reference design for multi-client agent translation (the translation apmw does); far too large to adopt. | no |
| VoltAgent/awesome-claude-code-subagents, davila7/claude-code-templates | 100+ agents each | Claude Code | No. Every installed agent description sits in context each turn; generic prompts. | no |
| BMAD-METHOD | persona-based skills method (PM, architect, dev, test) | Claude Code, Codex, any skills client | No. Heavy ceremony, large context. | no |
vercel-labs/skills (npx skills add) | skills-only distributor, ~80 clients | all | Skills only, which is all this repository ships; lockfile (skills-lock.json), check and update. | Distribution of this repository: npx skills add GravionLabs/skills, wrapped by /skills-manage for groups and pinned external packs |
| agentskills.io | the SKILL.md spec | all | Followed; client keys (context, disable-model-invocation) are allowed extensions. | yes |
| microsoft/apm | package manager: skills, agents, prompts, instructions, hooks, MCP; apm.yml dependencies with virtual paths and pins | Copilot, Claude, OpenCode, Cursor, Codex, Kiro, … | Superset (agents, hooks, MCP, dependencies); the extra primitives are no longer used here. | Distributed GravionLabs/apm-hub until its archive; replaced by npx skills since this repository ships skills only |
Evaluations
obra/superpowers v7.0.0 (desk analysis, 2026-10-12)
Why not adopt.
- No GitHub issue, sub-issue or release-branch model. Its unit of work is a spec file plus a plan file under
docs/superpowers/; this hub's unit is the Epic → Feature → PBI issue and a PR per PBI intorelease/next. Adopting it would mean rebuilding/deliveron top of it. - No stack skills. TDD is generic;
<lang>-implementationand<lang>-teststay necessary. - A SessionStart hook injects
using-superpowersinto every session and demands a skill check before any action, including clarifying questions. That collides with gated skills, and two mandatory process layers produce two review seats and two planning artifacts per change. - Cost: its own strict-cost spec reports ~13 USD and ~150 controller turns per plan run on Opus and proposes a mid-tier orchestrator subagent as the remedy. Work an issue by hand keeps one issue per session instead.
- A second distribution channel (a plugin per harness) next to
npx skills.
What we borrowed (Feature #113, MIT licensed): the report-file status contract, the review package as a file and a per-issue record for resuming. They lived in agents of apm-hub until its 3.0.0; delivery is now the deliver skill (delivery).
brainstorming and systematic-debugging have no counterpart here. They could be imported as standalone skills without the hook; that is an open owner decision, not yet an issue.
How to add an external pack
- Find the smallest installable unit: a SKILL.md directory or a plugin directory, not a repo root (
angular/skillsroot is not a package;dotnet/skillsroot installs everything). - Probe in a scratch directory:
npx skills add <owner/repo> --listand count theSKILL.mdfiles. More than ~5 descriptions per pack needs a reason. - Pin the ref (
<owner/repo>#<sha>) and add the source with its skill names to theexternalmap ofskills/delivery/skills-manage/scripts/Install-Skills.ps1, next to the group it belongs to. - Remove what the pack now covers from the own skill; keep project conventions and
## Review rules.