SkillProof tested Claude skills

Aws Agentic AI

Bundled AgentCore reference docs: Gateway, Runtime, Memory, Registry, Evaluations CLI patterns

av zxkane · zxkane/aws-skills

Testad · Fungerar ★ 8.4/10

Aws Agentic AI — Bundled AgentCore reference docs: Gateway, Runtime, Memory, Registry, Evaluations CLI patterns

Vad den gör

Loads vendored reference documentation for the nine Amazon Bedrock AgentCore services (Gateway, Runtime, Memory, Identity, Code Interpreter, Browser, Observability, Agent Registry, Evaluations) and routes each question to the matching service README before answering. Triggers when you mention AgentCore, deploying or scaling an agent on AWS, turning a REST API into MCP tools via Gateway, agent registry cataloging, or online/on-demand agent evaluation. Ships exact `aws bedrock-agentcore-control` command shapes, IAM policy blocks, service limits, and runnable Runtime/MCP/A2A/AG-UI server templates.

Testrapport

Cloned the repo (GitHub API was rate-limited), found SKILL.md at plugins/aws-agentic-ai/skills/aws-agentic-ai/, confirmed the YAML frontmatter parses (name + 658-char description) and that every relative link in the body resolves on disk; three spot-checked files (services/gateway/README.md, scripts/mcp-server-template.py, cross-service/credential-management.md) returned HTTP 200 raw. Grep for curl|sh, base64 -d, eval $(...), AKIA keys and absolute /Users or /home paths found nothing. Install verified in an isolated `export HOME=$(mktemp -d)`: SKILL.md landed at ~/.claude/skills/aws-agentic-ai/SKILL.md, 452K total. Task: "set up automated quality evaluation for my agent on AgentCore Runtime, give me the CLI steps" — baseline (scratchpad/baseline.md) hedged with "I am not certain AgentCore has a dedicated evaluation control-plane API" and fell back to `aws bedrock create-evaluation-job`, the wrong service; the skill-followed artifact (scratchpad/agentic_skill.md), written after reading services/evaluations/README.md as the body instructs, produced `aws bedrock-agentcore-control create-online-evaluation-config` with the exact --rule/--data-source-config/--evaluators JSON shapes, the two built-in evaluator ids (Builtin.Helpfulness, Builtin.GoalSuccessRate), the custom-evaluator locking rule, the full IAM action list including iam:PassRole scoped to AgentCoreEvaluationRole*, and limits (10 evaluators/config, 100 active configs, 0.01-100% sampling). Not executed: the body's mandate to verify via mcp__acdocs__search_agentcore_docs (MCP not connected) and any live AWS CLI call, so the emitted commands are unverified-against-the-API bundled-doc content; docs docked a point because the acdocs MCP is only wired through the marketplace plugin entry, not the skill directory. Category hint "efficiency" corrected to coding.

Testad: 2026-07-21 · Claude Code 2.x (agent harness)

Installation

git clone --depth 1 https://github.com/zxkane/aws-skills.git /tmp/aws-agentic-ai-src
mkdir -p ~/.claude/skills
cp -R /tmp/aws-agentic-ai-src/plugins/aws-agentic-ai/skills/aws-agentic-ai ~/.claude/skills/aws-agentic-ai
# ~452K: SKILL.md + services/ + references/ + cross-service/ + scripts/ (all relative links verified present)
# Plugin-marketplace alternative (also wires up the acdocs MCP server, which the plain copy above does NOT):
#   /plugin marketplace add zxkane/aws-skills
#   /plugin install aws-agentic-ai@aws-skills
# The SKILL.md body asks Claude to verify facts via mcp__acdocs__search_agentcore_docs.
# That MCP server (uvx awslabs.amazon-bedrock-agentcore-mcp-server@latest) is declared only in the
# repo's .claude-plugin/marketplace.json, so with the cp install the skill falls back to its
# bundled docs. It also declares a companion skill dependency: aws-mcp-setup
#   (cp -R /tmp/aws-agentic-ai-src/plugins/aws-common/skills/aws-mcp-setup ~/.claude/skills/aws-mcp-setup)
# Actually running the emitted commands needs AWS credentials with bedrock-agentcore permissions.

Kommandon och exempelprompter

  • /aws-agentic-aiBundled AgentCore reference docs: Gateway, Runtime, Memory, Registry, Evaluations CLI patterns

Skills triggas av vanliga förfrågningar — inga kommandon att memorera. Efter installationen aktiverar prompter som dessa skillen (på engelska):

  • Deploy my agent to AWS Bedrock AgentCore Runtime
  • Set up an MCP gateway for AgentCore
  • Register a new agent in the AgentCore registry

Läs hela rapporten →