aws-best-practice-research（AWS 最佳 practice 研究）

AWS Best Practice Research (with Optional Live Assessment)

Research and compile comprehensive best-practice checklists for any AWS service using the aws knowledge mcp server documentation search tools. Optionally assess live AWS resources against the compiled checklist.

Prerequisites

This skill requires the aws knowledge mcp server tools to be available:

• aws___search_documentation — search across AWS documentation topics
• aws___read_documentation — read full documentation pages
• aws___recommend — get related documentation recommendations

For the optional live assessment (Step 8):

• AWS CLI (aws) — must be configured with credentials that have read access to the target service
• jq — for parsing JSON output from AWS CLI commands

Workflow

Step 1: Identify Target Service and Assessment Scope

Determine from user input:

• AWS Service — e.g., ElastiCache Redis, RDS MySQL, MSK, EKS, Aurora, DynamoDB, etc.
• Focus areas — HA/DR, security, or all (default: all)
• Live assessment info (optional) — does the user provide any of the following?

- AWS credentials (environment variables, profile, or credential file path) - AWS Region (e.g., us-west-2) - Resource identifiers (cluster name, instance ID, table name, etc.)

If the service is ambiguous, ask the user to clarify (e.g., "RDS MySQL or RDS PostgreSQL?").

Record whether a live assessment is requested:

• If the user provides credentials + region + resource info → run live assessment after checklist
• If the user provides partial info → ask for the missing pieces before proceeding
• If the user provides no live resource info → skip live assessment, produce checklist only

Step 2: Sequential Documentation Search

Run the following 5 search queries one at a time, sequentially using aws___search_documentation. Do NOT run them in parallel — the aws knowledge mcp server has rate limits and parallel requests will trigger "Too many requests" errors.

Wait for each query to return results before sending the next one. Replace {SERVICE} with the actual service name (e.g., "ElastiCache Redis", "Amazon RDS MySQL", "Amazon MSK").

Query 1: "{SERVICE} best practices high availability disaster recovery"
  topics: ["general"]
  limit: 10

Query 2: "{SERVICE} Well-Architected reliability resilience best practices"
  topics: ["general"]
  limit: 10

Query 3: "{SERVICE} replication multi-AZ failover cluster mode backup"
  topics: ["reference_documentation", "troubleshooting"]
  limit: 10

Query 4: "{SERVICE} security encryption authentication access control"
  topics: ["general"]
  limit: 10

Query 5: "{SERVICE} Well-Architected security best practices"
  topics: ["general"]
  limit: 10

Rate limit protection: If any query returns a "Too many requests" error, wait 5 seconds and retry once. If it fails again, skip that query and continue with the next one.

Step 3: Read Key Documentation Pages

From the search results, identify and read the most important pages one at a time, sequentially using aws___read_documentation. Do NOT read multiple pages in parallel to avoid rate limiting. Prioritize these document types:

1. Well-Architected Lens pages for the service (Reliability, Security, Performance, Operational Excellence pillars)
2. Official best practices page
3. Resilience / disaster recovery page
4. Overall best practices page

Read each with max_length: 15000 to get comprehensive content. Typically 3-5 page reads are needed.

If a Well-Architected Lens exists for the service, it is the single most valuable source — always read it.

Step 4: Extract and Categorize Check Items

From all gathered documentation, extract individual check items and organize them into 5 mandatory categories (see references/output-template.md for the exact format):

Category 1: High Availability Architecture Items about: cluster mode, replication, replicas per shard, Multi-AZ, AZ distribution, node types, quorum.

Category 2: Disaster Recovery Items about: automatic/manual backups, retention policies, RPO/RTO documentation, Global Datastore / cross-region replication, failover testing, replication lag monitoring.

Category 3: Failover Planning Items about: Test Failover API, FIS resilience testing, client timeout/topology config, SNS event notifications, graceful degradation, WAIT command.

Category 4: Security Configuration Items about: encryption at-rest/in-transit, authentication (AUTH/RBAC/IAM), subnet groups, security groups, KMS keys, dangerous command renaming, RBAC metrics monitoring, IAM control plane policies.

Category 5: Others Items not covered by the above 4 categories, including but not limited to: auto minor version upgrade, engine version, node type selection (Graviton), CloudWatch monitoring, reserved memory, connection pooling, read routing, expensive commands, slow log, IaC management, Auto Scaling, cost tags, client retry logic, performance tuning, operational best practices.

Scope Boundary for Container / Orchestration Platforms

When the target service is a container or orchestration platform (EKS, ECS, Fargate, App Runner, Elastic Beanstalk), this skill focuses exclusively on the AWS infrastructure layer. All check items must be verifiable through AWS APIs (aws eks, aws ecs, aws ec2, aws iam, etc.).

Do NOT include check items that require kubectl, ECS Exec, or any in-cluster / in-task inspection to verify. These belong to a dedicated workload-level assessment skill.

For Amazon EKS, the infrastructure layer scope includes:

For Amazon ECS / Fargate, apply the same principle: check cluster, capacity providers, service auto-scaling, task definition registration, VPC configuration, and IAM roles — but do NOT check container-level health checks, resource limits, or task-internal configuration.

After generating the checklist, append a Scope Notice (see references/output-template.md for the exact format) directing users to a workload-level skill for the items that are out of scope.

For each check item, record:

• ID — category prefix + sequential number + priority suffix (e.g., HA-01-hi, DR-02-md, SEC-03-lo)

- Priority suffixes: -hi (High), -md (Medium), -lo (Low) - This embeds priority directly in the ID for quick visual scanning

• Check item name — concise, actionable
• Description — what to check and why, specific thresholds or values where applicable
• Source — which document/control it comes from (see source annotation rules below)
• Priority — High / Medium / Low (also kept as a separate column for filtering)

Step 5: Compile Source Annotations

Use consistent source tags throughout the checklist:

Step 6: Generate Output (Conditional)

The output depends on whether the user provided live assessment info in Step 1:

If NO live assessment info was provided → Generate Checklist Only

Generate the checklist content using the exact format defined in references/output-template.md, then write it to a local markdown file using the Write tool.

File naming: YYYY-mm-dd-HH-MM-SS-{SERVICE}-best-practice-checklist.md

• Replace YYYY-mm-dd-HH-MM-SS with the current timestamp (e.g., 2025-07-15-14-30-00)
• Replace {SERVICE} with a lowercase, hyphen-separated service name (e.g., elasticache-redis, amazon-eks)
• Example: 2025-07-15-14-30-00-elasticache-redis-best-practice-checklist.md
• Save the file in the current working directory

The checklist output must include:

1. Title with service name
2. One table per category (5 tables)
3. Source annotation legend
4. Key reference links section

After writing the file, inform the user of the file path.

If live assessment info WAS provided → Skip Checklist, Proceed to Step 8

Do NOT generate a separate checklist file. The assessment report (Step 8) will include the full checklist with assessment results in a single, comprehensive document. Generating both would be redundant.

Proceed directly to Step 8.

Step 7: Offer Next Steps (Checklist-Only Path)

This step only applies if you generated a checklist in Step 6 (no live assessment).

After writing the checklist file, suggest:

• "I can export this to a spreadsheet if you prefer."
• "If you provide AWS credentials and resource identifiers, I can assess a live resource against this checklist."

If the user provided live assessment info in Step 1, skip this step entirely — you should already be proceeding to Step 8.

Step 8: Live Resource Assessment (Optional)

Only execute this step if the user has provided credentials, region, and resource identifiers. If none were provided, skip this step entirely.

See references/assessment-workflow.md for the detailed per-service assessment procedure. The general flow is:

8.1 Prepare Environment

If the user provided a credential file path (e.g., env.sh), source it:

source <credential-file-path>

Verify access by running a simple describe command against the target service and region.

8.2 Collect Resource Configuration

Run the service-specific AWS CLI commands to gather the full configuration of the target resource. Execute independent commands in parallel to save time.

For ElastiCache Redis, the key commands are (see references/assessment-workflow.md for the full list):

• aws elasticache describe-replication-groups
• aws elasticache describe-cache-clusters --show-cache-node-info
• aws elasticache describe-cache-subnet-groups
• aws elasticache describe-cache-parameters
• aws elasticache list-tags-for-resource
• aws elasticache describe-snapshots
• aws elasticache describe-events

For other services, use the equivalent describe/list commands.

8.3 Map Configuration to Checklist

For each check item in the checklist, determine the assessment status:

For each item, record:

• The check ID and name (from the checklist)
• The assessment status (PASS / FAIL / WARN / N/A)
• A specific finding describing what was observed (include actual values)

8.4 Generate Assessment Report

Generate the assessment results using the format defined in references/output-template.md, then write it to a local markdown file using the Write tool.

This is the ONLY output file when a target resource is provided. The assessment report is self-contained and includes all checklist information (Description, Source, Priority) alongside the assessment results. Do NOT generate a separate checklist file.

File naming: YYYY-mm-dd-HH-MM-SS-{RESOURCE_ID}-assessment-report.md

• Replace YYYY-mm-dd-HH-MM-SS with the current timestamp (e.g., 2025-07-15-14-30-00)
• Replace {RESOURCE_ID} with the actual resource identifier, lowercase, hyphens for separators
• Example: 2025-07-15-14-30-00-my-redis-cluster-assessment-report.md
• Save the file in the current working directory

The report must include:

1. Resource Summary — key properties of the assessed resource (engine, version, node type, topology, etc.)
2. Assessment Results by Category — one table per category with full checklist columns (Check Item, Description, Source, Priority) PLUS assessment columns (Status, Finding)
3. Assessment Summary — counts of PASS/FAIL/WARN/N/A per category
4. Critical Issues — list of all FAIL items with Priority=High, with specific remediation guidance
5. Recommendations — grouped by urgency (Immediate / Short-term / Medium-term)
6. Source Annotations — legend for source abbreviations
7. Key Reference Links — documentation pages used

After writing the file, inform the user of the file path.

8.5 Offer Remediation

After presenting the assessment results, suggest:

• Which FAIL items can be fixed in-place (e.g., enabling backups, adding tags)
• Which FAIL items require resource recreation (e.g., encryption at rest)
• Whether you can help execute the remediation commands

Important Guidelines

• Be comprehensive: Search broadly, read deeply. The value of this skill is completeness. It's better to include a check item and mark it as lower priority than to miss it.
• Always cite sources: Every check item must have a source annotation. Users need to know where each recommendation comes from.
• Always use sequential requests: All searches and page reads must be executed one at a time, sequentially. Never send multiple aws knowledge mcp server requests in parallel. The MCP server has rate limits that will reject concurrent requests with "Too many requests" errors. Sequential execution is slower but reliable.
• Rate limit protection: If any MCP request returns a "Too many requests" error, wait 5 seconds and retry the same request once. If it fails a second time, skip that request and continue with the next step. Do not retry more than once per request.
• Focus on actionable items: Each check item should be something the user can verify against their actual configuration. Avoid vague recommendations.
• Include specific thresholds: When documentation specifies numbers (e.g., "at least 2 replicas", "reserved-memory-percent >= 25%"), include them in the check description.
• Note service-specific nuances: If a check only applies under certain conditions (e.g., "only if cluster mode enabled"), note that in the description.
• Live assessment is optional: Never fail or block if the user doesn't provide credentials. The checklist alone is a complete, valuable deliverable.
• Respect language: Always output in the same language as the user's conversation.