Everything you need to build effective AI agents — architecture, tools, prompts, and CLIs. Distilled from Anthropic, Google, and practitioner research into installable agent skills.
npx skills add tylergibbs1/agentsbestpractices
Pick the right architecture for your agent. 16 architecture patterns plus 6 dynamic workflow composition recipes, context engineering strategies, and a decision framework for when to go multi-agent.
| You'll learn | From |
|---|---|
| Single agent vs multi-agent decision framework | Anthropic's "Building effective agents" |
| Context engineering (progressive disclosure, offloading, caching) | Lance Martin's agent design patterns |
| Ralph loop for long-running tasks | Claude Code patterns |
| Long-running harness: initializer + worker agents, state files | Anthropic's "Effective Harnesses" |
| Coordinator, hierarchical, swarm patterns | Google's agent design guide |
| Dynamic workflow composition (fan-out-and-synthesize, adversarial verification, tournament, loop-until-done) | Claude Code dynamic workflows |
Build tools agents actually use correctly. Schema design, description writing, response formatting, and evaluation-driven improvement.
| You'll learn | From |
|---|---|
| Fewer, deeper tools beat many shallow tools | Anthropic's tool writing guide |
| 4-part description anatomy (what/when/how/constraints) | SWE-bench tool optimization |
| Response format control (concise vs detailed) | Slack MCP optimization |
| Programmatic tool calling (PTC) for token-efficient composition | Claude PTC announcement |
| Iterative tool design: evolving tools as models improve | "Seeing like an Agent" (Claude Code) |
| Evaluation loops with agent transcript analysis | Tool evaluation cookbook |
Write prompts that steer agents effectively. Model-agnostic principles with ready-to-use prompt blocks. Works across Claude, GPT, Gemini.
| You'll learn | From |
|---|---|
| System prompt structure with XML tags | Anthropic prompting best practices |
| Action vs suggestion steering | Claude 4.6 migration guide |
| Multi-window state management (tests.json, git checkpoints) | Long-horizon agent patterns |
| Copy-paste blocks for common fixes (verbosity, hallucinations, overengineering) | Production agent debugging |
Design CLIs that agents can use safely. Input hardening, schema introspection, context window discipline, and a retrofit checklist for existing CLIs.
| You'll learn | From |
|---|---|
| Raw JSON payloads over bespoke flags | Google Workspace CLI (gws) |
| Input hardening against hallucinations (path traversal, double encoding) | Agent-first CLI design |
| Schema introspection replaces documentation | Runtime discovery patterns |
| Multi-surface design (CLI + MCP + env vars) | Agent DX principles |
Each skill follows the Agent Skills spec:
skills/<skill-name>/
├── SKILL.md # Main instructions (loaded when triggered)
├── references/ # Detailed guides (loaded on demand)
└── evals/
└── evals.json # Evaluation tasks
Only the skill metadata is loaded into context at startup. The full SKILL.md is read when the skill triggers. Reference files are loaded only when needed — progressive disclosure keeps your context window clean.
These skills distill guidance from:
- Building effective agents — Anthropic
- How to write tools — Anthropic
- Prompting best practices — Anthropic
- Agent design patterns — Lance Martin
- Design pattern for your agentic AI system — Google Cloud
- You Need to Rewrite Your CLI for AI Agents — Justin Poehnelt
- Programmatic tool calling — Anthropic
- Seeing like an Agent: Lessons from Building Claude Code — Anthropic
- Effective Harnesses for Long-Running Agents — Anthropic
- Orchestrate subagents at scale with dynamic workflows — Anthropic
- Skill authoring best practices — Anthropic
- Fork the repo
- Create a branch (
feature/skill-name) - Follow the skill structure above — SKILL.md under 500 lines, references one level deep
- Include evals with realistic tasks (not trivial ones)
- Submit a PR
MIT