Agent Almanac

A library of executable skills, specialist agents, and pre-built teams for Claude Code and compatible AI tools. Define repeatable engineering procedures once and have AI agents execute them with built-in validation and error recovery. Compose specialists into review teams that catch issues a single reviewer would miss. Built on the Agent Skills open standard.

At a Glance

• 368 skills across 66 domains — structured, executable procedures
• 73 agents — specialized Claude Code personas covering development, review, compliance, and more
• 18 teams — predefined multi-agent compositions for complex workflows
• 34 guides — human-readable workflow, infrastructure, and reference documentation
• Interactive visualization — force-graph explorer with 368 R-generated skill icons and 9 color themes

How It Works

Ask Claude Code to review your R package, and the r-package-review team activates 4 agents — each following specialized skills for code quality, architecture, security, and best practices — then synthesizes their findings into a single report.

Works with

Skills follow the Agent Skills open standard and work with any tool that reads markdown:

Agents and teams use Claude Code's subagent architecture. For other tools, skills are the primary integration surface. See skills/README.md for setup instructions.

Quick Start

Choose a path. Full per-OS runbook with prereqs, verification, and updating: Installation guide.

Path 1 — Reference a skill (zero install)

> "Follow skills/commit-changes/SKILL.md to stage and commit my changes"

Path 2 — Claude Code plugin (recommended)

git clone https://github.com/pjt222/agent-almanac.git ~/dev/agent-almanac

# One-time local marketplace setup
mkdir -p ~/.claude-marketplace/{plugins,.claude-plugin}
cat > ~/.claude-marketplace/.claude-plugin/marketplace.json << 'EOF'
{
  "name": "local",
  "owner": { "name": "self" },
  "plugins": [{
    "name": "agent-almanac",
    "source": "./plugins/agent-almanac",
    "category": "development"
  }]
}
EOF
ln -s ~/dev/agent-almanac ~/.claude-marketplace/plugins/agent-almanac

claude plugin marketplace add ~/.claude-marketplace
claude plugin install agent-almanac@local

Auto-discovers all 368 skills and 73 agents. To use a team, read its definition in teams/<name>.md and spawn each listed member as a subagent via the Agent tool (subagent_type), coordinating them with SendMessage under the session's single implicit team. Windows / macOS variants in the Installation guide.

Path 3 — Global CLI (cross-framework)

npm install -g agent-almanac
agent-almanac detect              # show frameworks in cwd
agent-almanac install commit-changes

Reaches 12+ frameworks: Claude Code, Cursor, Codex, Gemini, Aider, OpenCode, Windsurf, Vibe, Hermes, OpenClaw. See cli/README.md.

Verify

cd ~/dev/agent-almanac
npm install && npm test           # repo integrity + README freshness
claude plugin list | grep agent-almanac

In a fresh Claude Code session, type /commit-changes — should resolve. Empty result → see Installation troubleshooting.

Explore visually

cd viz && npm install && bash build.sh && npm run dev
# Open http://localhost:5173 for the interactive force-graph explorer

Requires R 4.5.x or Docker; per-OS R paths in the Installation guide. See viz/README.md for build internals.

Directory Map

agent-almanac/
  .claude-plugin/  Plugin manifest for Claude Code plugin installation
  skills/          368 executable procedures across 66 domains
  agents/          73 specialist personas
  teams/           18 multi-agent compositions with 8 coordination patterns
  guides/          34 human-readable reference docs
  viz/             Interactive force-graph explorer with R-generated icons
  tests/           30 test scenarios for validation
  i18n/            Translations (10 locales: de, zh-CN, ja, es, caveman-lite, caveman, caveman-ultra, wenyan-lite, wenyan, wenyan-ultra)
  cli/             Universal installer CLI (npm install -g agent-almanac)
  scripts/         Build and CI automation
  sessions/        Tending session archives

Guides

New here? Start with Understanding the System. See all guides for the full categorized list.

Workflow

• Understanding the System — Entry point: what skills, agents, and teams are, how they compose, and how to invoke them
• Creating Skills — Authoring, evolving, and reviewing skills following the agentskills.io standard
• Creating Agents and Teams — Designing agent personas, composing teams, and choosing coordination patterns
• Creating Workflows — Authoring code-driven orchestration workflows — the meta contract, the agent/parallel/pipeline/phase primitives, and the capability rule
• Running a Code Review — Multi-agent code review using review teams for R packages and web projects
• Managing a Scrum Sprint — Running Scrum sprints with the scrum-team: planning, dailies, review, and retro
• Visualizing Workflows with putior — End-to-end putior workflow visualization from annotation to themed Mermaid diagrams
• Running Tending — AI meta-cognitive tending sessions with the tending team
• Running a Translation Campaign — End-to-end guide for translating all skills, agents, teams, and guides into supported locales using the translation-campaign team
• Unleash the Agents — Structured multi-agent consultation at three tiers for open-ended hypothesis generation
• Team Assembly Prompt Patterns — How to phrase requests to Claude Code for multi-agent team work at every level of specificity
• Production Coordination Patterns — Real-world multi-agent orchestration patterns: barrier synchronization, silence budgets, health checks, degraded-wave policies, and cost-aware scheduling
• AgentSkills Alignment — Standards compliance audits using the agentskills-alignment team for format validation, spec drift detection, and registry integrity
• Edge Computing Deployment — Install agent-almanac skills on edge LLMs (Gemma 4 via AI Edge Gallery) with distilled content, token budgets, and offline bundles
• Self-Continuation Loops Playbook — Choose among ScheduleWakeup, CronCreate loops, and loop.md; select sentinels; plan for the 7-day age-out
• Headless WebGL Verification — Proving a WebGL/GPGPU/WebAudio app rendered in headless Chromium — SwiftShader/ANGLE flags, capability probes, pixel-luminance assertions, HMR traps
• Driving the Copilot Review Loop — Converging a PR to a clean GitHub Copilot review pass — thread IDs, the fix/reply/resolve/re-request cascade, and reading the bot's verdict

Infrastructure

• Installation — OS-aware install runbook covering plugin install, global CLI, prereqs, verification, and updating across Linux, macOS, Windows, WSL2, and Codespaces
• Setting Up Your Environment — WSL2 setup, shell config, MCP server integration, and Claude Code configuration
• Symlink Architecture — How symlinks enable multi-project discovery of skills, agents, and teams through Claude Code
• R Package Development — Package structure, testing, CRAN submission, pkgdown deployment, and renv management
• WSL Maintenance & Claude Code Reference — WSL2 vhdx disk reclamation, Claude Code permission modes, and periodic security-scan greps for a WSL-based dev environment
• Protecting GitHub Repositories — Honest threat model, tiered checklist, rulesets vs branch protection, and the CI-bot bypass problem for hardening a public GitHub repo

Reference

• Quick Reference — Command cheat sheet for agents, skills, teams, Git, R, and shell operations
• Agent Best Practices — Design principles, quality assurance, and maintenance guidelines for writing effective agents
• Agent Configuration Schema — YAML frontmatter field definitions, validation rules, and JSON Schema for agent files
• Content Styleguide — Canonical markdown formatting conventions for skills, agents, teams, and guides — tables, code fences, headings, lists, and links
• Choosing an HTR/OCR Engine — Decision framework for selecting a handwritten-text-recognition engine — deployment model, API entitlement, vision-LLM vs dedicated HTR, CER pitfalls
• The Caveman Spellbook — Six grunt-level compression modes for agent-almanac content — a homage to JuliusBrussee/caveman, from lite filler-stripping to extreme classical Chinese abbreviation

Design

• Extracting Project Essence — Multi-perspective framework for extracting skills, agents, and teams from any codebase using the metal skill
• Epigenetics-Inspired Activation Control — Runtime activation profiles controlling which agents, skills, and teams are expressed, grounded in molecular epigenetics
• Understanding the Synoptic Mind — The adaptic concept — panoramic synthesis through simultaneous multi-domain awareness, theoretical foundations, and practical use
• Agent Memory Hygiene — Three-layer model — weights, retrieval, behavior — for diagnosing what kind of forgetting a memory problem actually needs and applying the right tool

Translations

See i18n/README.md for the translation contributor guide.

Plugin Packaging

Agent-almanac is packaged as a Claude Code plugin at .claude-plugin/plugin.json. When installed, Claude Code auto-discovers all skills and agents:

Teams are not a plugin-native content type — activate a team by reading teams/<name>.md and spawning each listed member as a subagent via the Agent tool (subagent_type), coordinating them with SendMessage under the session's single implicit team (see Creating Agents and Teams). TeamCreate is a gated FleetView/cloud fallback, not the path for ordinary interactive sessions.

For step-by-step plugin install (POSIX + Windows + macOS variants, prereqs, verification, troubleshooting), see the Installation guide.

Contributing

Contributions welcome! Each content type has its own guide:

• Skills — skills/README.md for format and consumption
• Agents — agents/README.md for template and best practices
• Teams — teams/README.md for coordination patterns
• Guides — guides/README.md for categories and template

Update the relevant _registry.yml when adding content, then run npm run update-readmes.

Support

If Agent Almanac makes your AI tools more capable, consider sponsoring its development.

Reliable AI assistance requires structured knowledge — and maintaining that structure is work that the models themselves cannot do.

License

MIT License. See LICENSE for details.