Skip to content

Repository files navigation

fable-pack: curated plugin marketplace for Claude Code

CI MIT license 83 skills 23 subagents no executable code

A curated, license-clean plugin marketplace for Claude Code. We reviewed 1,000+ skills and agents from the most popular community packs and kept fewer than 10% — every survivor either contains knowledge a frontier model would otherwise hallucinate, or a battle-tested methodology with hard rules.

/plugin marketplace add Mixard/fable-pack

What's new

Date Release Highlights
2026-07-21 fable-workflows 0.2.0 project-cartography — living three-file project map (CODEMAP / PROJECT_STATE / DECISIONS) so large projects survive session boundaries; lifecycle skills now chain
2026-07-21 fable-guard 0.2.0 Stop hook: one-shot stale-map reminder in cartography-mapped projects
2026-07-21 fable-guard 0.1.0 New opt-in plugin: PreToolUse hooks that deterministically block secret leaks and curl | sh before execution
2026-07-21 fable-knowledge 0.3.0 n8n-selfhosted-ops — first original skill (CLI import without API key, systemd env, Telegram HITL without broken sendAndWait)
2026-07-21 fable-agents 0.2.0 Explicit model tier on all 23 agents: 7 opus / 15 sonnet / 1 haiku — no agent inherits the expensive orchestrator model
2026-07-21 repo Validator now enforces agent model tiers and link integrity; monthly freshness sweep re-verifies version-fragile skills

Full history in CHANGELOG.md.

Why this selection

Most community packs were written for weaker models: they teach the model to write React, name variables, or "think step by step." A frontier model does not need any of that — shipping it as a skill just burns context tokens on things the model already does well.

What a frontier model still gets wrong is narrow and specific:

  • Exact API schemas and CLI flags — it will confidently invent a plausible-but-wrong parameter name
  • Version-specific behavior after its knowledge cutoff — Next.js 16, Swift 6.2, iOS 26, Angular 21, n8n 2.8
  • Platform gotchas that contradict intuition — the documented behavior that no amount of reasoning predicts
  • Methodologies with hard numeric rules — cold-email benchmarks, A/B sample sizes, TDD's iron law

That is the entire selection bar. Of 269 skills triaged from ECC alone, 18 survived. Of ~745 agents in wshobson/agents, 23 made it in.

Why it fits Fable

The pack is tuned for a frontier orchestrator model (Claude Fable / Opus) running the main session:

  • Skills carry facts, not lectures — short, dense, no process rituals. The orchestrator reads exact schemas and flags instead of re-deriving or hallucinating them.
  • Every agent declares the cheapest model that does the job — 7 on opus (architecture, code review, security), 15 on sonnet (implementation and operations), 1 on haiku. No agent silently inherits the expensive orchestrator model. The orchestrator delegates mechanical skill work down-tier and keeps judgment work for itself.
  • Nothing competes with the model — no meta-frameworks, no personas, no "orchestration systems" that fight the harness. The pack only fills gaps.

Install

Install any subset — plugins are independent:

/plugin install fable-knowledge@fable-pack
/plugin install fable-agents@fable-pack
/plugin install fable-workflows@fable-pack
/plugin install fable-marketing@fable-pack
/plugin install fable-guard@fable-pack

Plugins

fable-knowledge — 48 skills

Knowledge-only: exact API schemas, CLI flags, version-specific behavior, platform gotchas. No process rituals.

Category Skills
Media / video fal-ai-media, videodb, remotion, ffmpeg-media-recipes, manim-explainers, playwright-demo-videos, html-slides, ios-icon-gen
API integrations x-api, jira-integration, nutrient-api, free-tier-scraper-apis, mcp-server-configs, mailtrap-email-integration, laravel-plugin-discovery, codehealth-mcp, claude-devfleet, agent-payment-x402
Web / frameworks bun-runtime, nuxt4-patterns, nextjs-turbopack, react-performance, angular-developer, wcag22-reference, pm2-node-services
Data clickhouse, database-migrations, postgres-tips, prisma-patterns, regex-llm-hybrid
Scientific APIs pubmed-database, uspto-database, gget
Apple (2025+ APIs) swift-concurrency-6-2, ios26-liquid-glass, apple-foundation-models
Language niches kotlin-exposed, kotlin-ktor, perl-modern, cpp-core-guidelines, tinystruct-patterns
Packaging / ops nuitka-windows-packaging, flox-environments, uncloud, windows-desktop-e2e, n8n-selfhosted-ops
EVM / DeFi evm-gotchas, defi-amm-security

fable-agents — 23 subagents

Deep specialist subagents with concrete, tool-specific knowledge.

Category Agents
Languages python-pro, rust-pro, golang-pro, java-pro, sql-pro, bash-pro
Review / security code-reviewer, architect-review, security-auditor
Infrastructure kubernetes-architect, terraform-specialist, cloud-architect, deployment-engineer, database-admin, database-architect, database-optimizer
Reliability devops-troubleshooter, incident-responder, observability-engineer, performance-engineer
Product frontend-developer, backend-architect, test-automator

Every agent declares an explicit model tier — opus only where judgment is the product, sonnet for implementation, haiku for mechanical work.

fable-workflows — 15 skills

Battle-tested methodologies with hard rules, adapted from obra/superpowers:

test-driven-development, systematic-debugging, brainstorming, writing-plans, executing-plans, verification-before-completion, using-git-worktrees, subagent-driven-development, requesting-code-review, receiving-code-review, finishing-a-development-branch, dispatching-parallel-agents, project-cartography, getting-unstuck, critical-review

Two critical-thinking skills complement each other: getting-unstuck fires when a path is declared impossible and breaks the wall with tested hypotheses; critical-review fires when nothing is visibly wrong and hunts blind spots — stale assumptions, data contradictions, dismissed alternatives — through a mandatory 7-lens pass. Blocked opportunities found by the review route into getting-unstuck.

The lifecycle skills chain: brainstorming settles the design, writing-plans turns it into phased tasks, executing-plans runs them, verification closes them out — and project-cartography keeps a three-file living map (CODEMAP, PROJECT_STATE, DECISIONS) so on large projects every new session starts from ~300 lines instead of re-reading the codebase. The maps serve the model, never command it: what counts as substantive stays a judgment call.

fable-marketing — 20 skills

Marketing frameworks with concrete numbers, benchmarks, and templates.

Category Skills
Acquisition cold-email, ads, ad-creative, prospecting, sms, directory-submissions
SEO seo-audit, ai-seo, programmatic-seo, aso, competitors
Conversion / retention ab-testing, popups, offers, churn-prevention, pricing
Foundations product-marketing, customer-research, copy-editing, revops

fable-guard — deterministic guardrails (opt-in)

PreToolUse hooks that block secret patterns (API keys, tokens, private keys) in shell commands and file writes, plus dangerous shell patterns (curl | sh, --dangerously-skip-permissions) — before they execute. A model can be tricked or forget; a hook fires every time. Also ships a Stop hook that reminds — exactly once, and only in projects keeping a CODEMAP.md — when code changed but the project map did not (pairs with the project-cartography skill; whether the change was substantive stays the model's call). Dependency-free Python, fully readable in plugins/fable-guard/hooks/.

MCP servers

fable-pack ships no MCP servers — that is a deliberate part of the security model below, not a gap. Bundled servers auto-start with your session and widen the attack surface; a knowledge pack has no business running processes.

Instead, the mcp-server-configs skill (fable-knowledge) carries exact, pinned launch configs for the servers people actually use — Jira, GitHub, Supabase, Playwright, fal.ai, Cloudflare, Vercel, and others — so the model writes a correct .mcp.json on the first try and you decide what runs. Related skills cover the MCP surface where precision matters: claude-devfleet (exact tool signatures for parallel-agent orchestration), codehealth-mcp, laravel-plugin-discovery, and nutrient-api.

Security model

The four content plugins are inert by design: markdown only — no hooks, no MCP servers, no code that runs on install or load. Reference scripts inside skills are examples the model may run with your normal tool permissions, never automatically. Installing them cannot send your code or keys anywhere.

fable-guard is the single deliberate exception: it ships PreToolUse hooks, which Claude Code runs automatically once the plugin is installed. That is its entire purpose — deterministic enforcement the model cannot skip. It is opt-in, offline, stateless, and small enough to audit in one sitting before installing.

Selection principle

Kept only if a strong model would otherwise get it wrong:

  • exact API schemas, endpoints, request/response shapes
  • exact CLI flags and commands
  • version-specific behavior and recent platform changes (2025+)
  • platform gotchas that contradict intuition
  • curated reference data and working non-trivial scripts
  • methodologies with hard rules and measured numbers

Dropped: generic best practices, personas, thin wrappers, pack meta-tooling, and anything a frontier model produces reliably on its own.

Conventions

  • Skills: plugins/<plugin>/skills/<name>/SKILL.md, two-field frontmatter (name, description), optional references/ and scripts/.
  • Agents: plugins/fable-agents/agents/<name>.md with name, description, and an explicit model tier.
  • Model policy: every agent declares the cheapest model that does the job well — opus only for judgment-heavy work (architecture, code review, security), sonnet for implementation and operations, haiku for mechanical tasks. Agents never inherit the orchestrator's model. Skills run inline; when a skill implies substantial mechanical work, the orchestrator should delegate it to a subagent on a lower tier.
  • English only, no emojis, SKILL.md under 800 lines.
  • python3 scripts/validate.py checks all of the above (including required agent model tiers and link integrity); CI runs it on every PR.

Contributing

Missing a skill? Found stale facts? Open an issue — there are templates for content requests and outdated information reports. Outdated-info reports are the most valuable signal for a knowledge pack: version-fragile skills are re-verified against current releases on a monthly sweep. See CONTRIBUTING.md for the bar and format.

Sources and licensing

This pack is MIT licensed. Parts are adapted from permissively licensed community packs — see ATTRIBUTIONS.md for the full mapping: affaan-m/ECC, wshobson/agents, obra/superpowers, coreyhaines31/marketingskills. Full change history in CHANGELOG.md.

About

Curated plugin marketplace for Claude Code: 80 knowledge skills + 23 subagents, <10% keep rate from 1000+ reviewed. Markdown-only, no executable code.

Topics

Resources

Contributing

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages