Batch 14 Notes — Skill Marketplaces + Plugin Packs

Batch theme: Skill marketplaces + plugin packs (subagent collections / large catalogs) Frameworks analyzed: 10 Date: 2026-05-26

Roster Table

Catalog Size Comparison

Intra-Batch Patterns

Pattern 1 — Multi-harness portability is diverging: Two distinct approaches emerged. wshobson-agents uses a Python adapter that generates harness-native artifacts at build time (single Markdown source → 5 runtime outputs). alirezarezvani-claude-skills takes BYO-sync: all harness formats are pre-committed to the repo so users clone and copy what they need. Both approaches solve the same problem (one codebase, multiple tools) but make opposite tradeoffs — adapter-generated artifacts stay in sync automatically; BYO-sync gives zero runtime dependencies. VoltAgent took the simplest path: one repo per harness (claude-code-subagents + codex-subagents as explicit mirrors).

Pattern 2 — Format fragmentation within "skills": Across the 10 repos, the word "skill" means five different things: (a) Claude Code SKILL.md plugin primitive (wshobson, daymade, jeffallan); (b) flat Markdown behavioral prompts in a skills/ folder (alirezarezvani); (c) factory-droid droids labeled "skills" (droid-tings); (d) any behavioral instruction file regardless of front matter (jeremylongshore); (e) the entire plugin catalog called "skills" (jeffallan, daymade). No two repos agree on the precise definition.

Pattern 3 — Catalog scale inversely predicts methodological depth: The three repos with the highest component counts (jeremylongshore: 3,010+, droid-tings: 530, alirezarezvani: 441) have the weakest per-component documentation standards. The three with the strongest methodology (daymade, jeffallan, wshobson-agents) have the smallest catalogs by item count. The exception is wshobson-agents, which maintains rigorous standards at scale through automation (plugin-eval certification, adapter generation).

Pattern 4 — Orchestration falls into two camps: Either a repo actively uses Claude Code's Task tool to spawn typed subagents (wshobson-commands, wshobson-agents, qdhenry, jeremylongshore), or it documents chains without automating them (jeffallan, alirezarezvani, daymade). There is no middle ground — repos either commit to subagent spawning architecturally or leave chaining entirely to the user.

Pattern 5 — Deprecation-within-batch: wshobson-commands and wshobson-agents are the same author's successive generations. The commands repo README explicitly declares itself superseded. This is the only case in the batch where a single author's evolution is visible across two frameworks, providing a clear trajectory: flat slash commands → plugin marketplace with multi-harness adapter.

Most Interesting Finds

wshobson-agents: The multi-harness Python adapter is the most architecturally ambitious approach in the batch. The core insight — "a skill written once should produce idiomatic artifacts for every tool, not lowest-common-denominator translations" — is executed via a three-layer adapter stack (harness adapter → context template → output formatter) with plugin-eval certification (static checks, LLM quality judge, Monte Carlo variance testing) before any skill goes to the marketplace. The four-tier model routing per agent (Opus/Sonnet/Haiku/inherit in frontmatter) with explicit override capability at install time is a capability no other catalog implements.

jeffallan-claude-skills: The application of EARS (Easy Approach to Requirements Syntax) — a formal requirements notation from aerospace/defense — to AI-assisted feature specification writing is unique across the entire batch. Combined with the the-fool adversarial review skill and spec-miner retrospective spec extraction (deriving specs from existing code), this is the only catalog treating requirements engineering as a first-class concern. The common-ground command's git-remote-based project fingerprinting and Mermaid assumption dependency graphs (--graph mode) are similarly without parallel.

Tier C Repos

droid-tings (44 stars): Written as Tier C due to minimal community adoption (44 stars), no versioning, unclear maintenance status (no recent pushes found in metadata), and thin per-component documentation. The Factory Droid / GenerateDroid tool integration is architecturally novel but not usable without Factory platform access which is external and undocumented in the repo itself. As a pure catalog exercise the skill count (375) is high, but most skills are brief and lack the structural depth seen in Tier A/B entries.

Cross-References Discovered

• wshobson-commands is an explicit predecessor to wshobson-agents: The commands README says "this repo has been superseded by the agents plugin marketplace." The commands repo is preserved for users who prefer the slash-command pattern without the plugin architecture.

• voltagent-subagents and voltagent-codex-subagents are runtime mirrors: Same organizational structure, same category taxonomy (01-core-development through 07-specialized), same 8–9 agents per category. One is .md for claude-code, one is .toml for codex-cli. They diverge only on format primitives.

• daymade is an explicit Anthropic official skill-creator fork: The skill-creator SKILL.md names Anthropic's original as inspiration and competitor, documents what it improves (AskUserQuestion structure, security gate, description optimizer). This is the clearest "upstream-downstream" relationship in the batch.

• alirezarezvani's karpathy-coder plugin: The karpathy-coder plugin bundles Andrej Karpathy's four coding principles (Think Before Coding, Simplicity First, Surgical Changes, Goal-Driven Execution) as a karpathy-reviewer.md agent. This is the only framework in the batch where a named external practitioner's methodology is packaged as a first-class plugin.

• jeffallan's Atlassian workflow dependency: The 9 workflow commands in jeffallan require an external Atlassian MCP server not bundled with the repo. Without it, ~9/12 commands are degraded. This makes jeffallan's effective full feature set dependent on an external enterprise subscription.

• EARS format connection: jeffallan's feature-forge skill uses EARS (Easy Approach to Requirements Syntax), a notation standard developed by Alistair Mavin at Rolls-Royce for safety-critical systems. No other repo in this batch or in the seed set uses EARS. The specification template in references/ears-syntax.md is a genuine bridge between formal requirements engineering and AI-assisted development.