Looking Back on AI Utilization by One AWS Engineer #devio2026 Presentation Materials
This page has been translated by machine translation. View original
Presented at DevelopersIO 2026 Osaka held on 2026/09/28 under the title "Looking Back on AI Utilization as One AWS Engineer".
Thank you to everyone who attended!
I will share the slides and content from the presentation on this blog.
Slides
What follows is an introduction to the slide content.
Intro
What I Want to Convey in This Session
Over the past year or two, the way we use generative AI has changed significantly.
- The tools we use have been replaced in short cycles
- Our ways of working and mindset have changed along with that
I will reflect on the changes experienced by one AWS engineer while introducing the AI tools I have actually been using.
Main Tools to Introduce
- Various AI tools
- simonw/llm, etc.
- Claude Code (and useful tools used with it)
- AWS MCP Server
- DuckDB Skills, etc.
Before the Claude Code Era
How things were used around the first half of 2025
How Things Were Used Around the First Half of 2025
| Tool | Where Used | Purpose |
|---|---|---|
| AI-Starter | Browser | AI chat |
| simonw/llm | Terminal | One-off queries |
| Claude Code | Terminal | Verification only |
AI-Starter: Generative AI Chat Environment
AI-Starter: Main Use Cases
- Having it write shell scripts and simple Lambda functions
- Having it summarize Terraform and CDK diffs, etc.
simonw/llm: Running LLMs from the Command Line
- Supports various models including OpenAI and Bedrock
- Can pass standard input via pipes
- Rich set of extensions including plugins and templates
simonw/llm: Introductory Blog Post
simonw/llm: Main Use Cases
- Having it summarize and explain CloudFormation and Terraform code
- Having it summarize and explain AWS CLI command output and help, etc.
As for Claude Code…
- At the time (around 2025/03), it was still in preview
- I was trying to use it by calling Claude 3.7 Sonnet from Bedrock
- I could not use it casually since it was pay-as-you-go
However, once it went GA and Claude Code became available under Pro and Max (subscription plans), my usage accelerated rapidly.
The Claude Code Era
From the second half of 2025 to the present
Claude Code Utilization for an AWS Engineer
| Use Case | Main Tools |
|---|---|
| Technical research / Resource investigation | AWS MCP Server |
| Log and cost analysis | Skills (DuckDB Skills) |
| Writing IaC (Terraform) | Various |
What is AWS MCP Server
It is a managed MCP server provided by AWS and is part of Agent Toolkit for AWS.
- Can perform documentation searches and retrieve service information
- With IAM authentication, scripts that call APIs in a sandbox environment can also be executed
Note: Difference Between AWS MCP Server and awslabs/mcp
awslabs/mcp is a collection of MCP servers that AWS has published as OSS. Agent Toolkit for AWS is the successor to awslabs/mcp.
AWS MCP Server: Main Use Cases
- Looking up AWS specifications and new features from the latest documentation
- Checking the status and configuration of resources in an AWS environment
What is DuckDB Skills
- DuckDB is a database system specialized for online analytical processing (OLAP)
- DuckDB Skills is an official collection of DuckDB skills for Claude Code
- Allows querying CSV, Parquet, and JSON in natural language for analysis and aggregation
Excerpt of Available Skills
| Skill Name | Brief Overview |
|---|---|
| read-file | Automatically detects file type from extension for CSV/JSON/Parquet etc. and reads the file and confirms schema |
| query | Executes SQL/natural language queries against attached DBs or files |
| read-memories | Searches past Claude Code session logs to surface decisions and TODOs |
| duckdb-docs | Full-text searches the DuckDB/DuckLake official documentation and blog |
Having It Write IaC (Terraform)
- The gut feeling is that from Opus 4.6 onward, even in a plain state, it generally produces what you expect
- If using it as a team, having the surrounding environment in order is convenient
- Formatter/linter: terraform fmt, tflint
- Security scanning: trivy
- Incorporating these into pre-commit or CI
- Tools I am interested in but have not been able to use much
- terraform-mcp-server : Can reference the latest information from Terraform Registry
- Terraform Skills : HashiCorp official Skills for Terraform
Helpful Blog Posts
Toward the Conclusion
Summary of Tool Transitions
| Tool | 2025 First Half | 2025 Second Half | ~ Present |
|---|---|---|---|
| AI-Starter | High | Medium | Low |
| simonw/llm | High | Medium | - |
| Claude Code | Low | High | Essential |
Sidebar: Tools Our Team is Using #1
- Gemini (Notebook, etc.)
- Compile sources into Notebooks and use for catch-up and retrospectives. Easy to share with the team too
- Search and explain data on Google Drive
- Slack AI
- Casually query internal information before asking someone
Sidebar: Tools Our Team is Using #2
- Kiro
- An AI agent tool made by AWS, available as both an IDE and CLI
- Can integrate with IAM Identity Center, keeping license management entirely within AWS
- As CCoE, we are rolling out Kiro's utilization infrastructure to user departments
- AWS DevOps Agent
- A managed AI operations agent provided by AWS
- Makes querying AWS resources easy
- Can automatically run alert investigations, reducing the burden of initial investigation
- Has few prerequisites and setup requirements, making it easy to recommend for adoption
Conclusion
I looked back on the transitions among the AI tools I have been using.
As AI tools keep getting more convenient, I feel the need to change my own thinking and approach as well.
- Being more aware of and involved in higher-level layers
- Extracting what I am thinking from my head and handing it to AI
It seems that verbalization and communication will become even more important going forward.
