AWS Bedrock AgentCore Skill
A Claude Code plugin (and agent skill) that puts the best practices for building AI agents on AWS, with Amazon Bedrock AgentCore at the center, in one place. Instead of sending the coding agent to search across dozens of AWS docs or work things out by trial and error, it hands over a consolidated, official, source-cited playbook so the agent goes straight to the right approach and can show you the source behind every recommendation.
Scope: Strands Agents, Amazon Bedrock (Converse, Guardrails, Knowledge Bases), and Amazon Bedrock AgentCore (Runtime, Memory, Gateway, Identity, Browser/Code Interpreter), plus Terraform-first IaC and CloudWatch/OpenTelemetry observability.
Built and verified with Claude Code multi-agent workflows: ~140 subagents, ~15M tokens, and 800+ official-documentation reads went into researching, writing, and adversarially QA-ing this skill.
What it is
It is a Distilled Knowledge Skill (DKS): an entire domain (Strands Agents, Amazon Bedrock, Bedrock AgentCore) reduced to its executable essence, and kept current as the surface shifts month to month. The research is already done; you skip straight to building.
Concretely, it is not a single template or a code generator. It is a routing layer plus a reference library:
- a
SKILL.md that acts as a decision tree: it maps the user's use case to a recommended stack and to the exact reference files to open;
- 20 reference files (~19,000 lines) covering each area in depth, with an inline
_Source: URL on every best practice and code snippet;
- assets: a service-selection matrix, a model-selection guide, a pre-production checklist, ready-to-adapt IAM policies, and copy-paste starter snippets;
- a central source index (
sources.md) mapping each topic to its official URL (369 unique official sources).
The agent loads only the files a task needs (progressive disclosure), so it stays useful without pulling the whole library into context.
Why use it
Building agents on AWS (especially Bedrock AgentCore) means a lot of scattered documentation and a fast-changing API surface. Left on its own, a coding agent either crawls across many pages or proceeds by trial and error, and still gets the version-specific details wrong. This skill removes both problems:
- The best practices are already gathered and organized. The agent does not have to research half the internet: the relevant official guidance for each area is collected in one place and routed by use case, so it goes straight to the right approach instead of probing around.
- It is current and source-cited. Bedrock AgentCore is recent and changes often. The skill encodes today's official answers and attaches the documentation URL to each one, so the agent (and you) can verify a recommendation instead of trusting it blindly. There are 636 inline source citations across the reference files.
- It prevents the common, non-obvious mistakes. These are the kind of errors that look correct in review and only fail at deploy: using the legacy
InvokeModel instead of the Converse API, passing serviceTier as a bare string, calling a deprecated structured_output(), setting a 1-hour prompt-cache TTL on a model that only supports 5 minutes, ignoring the ARM64 AgentCore runtime contract, or mis-sizing max_tokens and hitting the 5x token burndown. The skill documents the correct form for each.
- It picks the right pattern, not just the API. The decision tree distinguishes a simple chatbot from a tool-using agent, RAG, a multi-agent system, a serverless production deployment, and so on, and points to the matching reference. It also covers when a managed alternative (Bedrock Agents, Flows, the Responses API) fits better than the code-first path.
- It is maturity-aware. Every feature is labelled GA or Preview, and Preview is never recommended as a production default.
- It does not hard-code values that rot. Live prices, default quotas, and current model IDs are deferred to the AWS console and model cards; the skill teaches the durable shape (how caching is billed, how burndown works) and points to the live source for the exact numbers.
Quickstart
/plugin marketplace add ferdinandobons/AWSBedrockAgentCoreSkill
/plugin install aws-bedrock-agentcore-skill@aws-agent-skills
Then describe the agent you want to build. The skill triggers automatically (see How it triggers). Full install options, including local testing, are below.
How it triggers
