Introduction
Hello everyone!
I recently participated in the AWS Summit Japan 2026 AI-DLC Hackathon, which was held alongside AWS Summit Japan 2026.
The theme of this hackathon was extremely challenging: build a service that “makes people useless,” while following the AI-DLC approach!
https://pages.awscloud.com/summit-japan-2026-hackathon-reg.html
It was a highly competitive event with 117 teams entering, but in the end, we were able to win 3rd place!
https://x.com/haruki_web3/status/2070433400918553035
Product GitHub Repository
https://github.com/mashharuki/AWS-SummitHackathon-2026
In this technical blog post, I would like to share the things we did to make the most of AI-DLC and maximize its benefits throughout the hackathon.
What Is AI-DLC?
Overview of AI-DLC
AI-DLC stands for AI-Driven Development Lifecycle, a new development methodology proposed by AWS for the AI era.
It is a software development approach that restructures the entire development lifecycle around a division of roles where AI executes and humans supervise.
The details are explained in the following article and GitHub repository, so I will only briefly summarize the concept here.
https://zenn.dev/aws_japan/articles/aidlc-workflows
https://github.com/awslabs/aidlc-workflows/tree/main
The AI-DLC workflow is broadly divided into three phases:
- Inception phase, which covers requirements definition and design
- Construction phase, which covers implementation
- Operation phase, which covers operations and deployment
How Is It Different from cc-sdd?
Another well-known approach for practicing spec-driven development is cc-sdd.
https://github.com/gotalab/cc-sdd
At a high level, the overall flow of cc-sdd and AI-DLC is similar.
Both generally follow the process of requirements definition → design → implementation, with differences in the finer details.
After actually using AI-DLC extensively, I felt that AI-DLC is designed with practical enterprise use cases very much in mind.
The instructions given to AI and summaries of the outputs generated as a result are recorded in great detail in a file called audit.md.
https://github.com/mashharuki/AWS-SummitHackathon-2026/blob/main/aidlc-docs/audit.md
This makes it possible to look back later and understand who collaborated with AI, at what timing, and based on what kind of thinking during the development of the product. I found this aspect extremely compelling.
This audit.md file is something cc-sdd does not provide by default. I also felt that the volume and richness of the documents generated during the Inception phase, such as requirements and design documents, were far more comprehensive compared to cc-sdd.
https://github.com/mashharuki/AWS-SummitHackathon-2026/tree/main/aidlc-docs
Another difference is the Operation phase. AI-DLC generates documents that even cover the deployment of the product you build.
For example, when using AWS CDK, it carefully summarizes the CDK stack for the product, setup procedures, and deployment commands.
You can do something similar with cc-sdd as well, but in that case, you need to explicitly instruct the AI to do so.
What We Did in the AI-DLC Hackathon
The standard features of AI-DLC are as described above.
From here, I will summarize the specific things we did during the hackathon.
The Quality of Requirements Documents and Shared Goal Alignment Are Crucial
The first thing our team focused on was repeatedly running the Inception phase and aligning the goal image among team members.
Unlike cc-sdd, one of the strengths of AI-DLC is that it generates a rich set of documents during the requirements definition and design phase, making it easy for the team to review them together.
Before advancing to the first screening round, we repeated the process of creating, reviewing, and updating the documents more than four or five times, thoroughly improving their quality toward our final goal.
Creating Subagents and Skills for AI-DLC
To practice AI-DLC effectively, we prepared and tested subagents and Skills in advance.
https://github.com/mashharuki/AWS-SummitHackathon-2026/blob/main/.claude/agents/aidlc-specialist.md
https://github.com/mashharuki/AWS-SummitHackathon-2026/tree/main/.claude/skills/aidlc-workflow-guide
We also created a dedicated review Skill that summarized the hackathon overview and judging criteria, as well as a strategy Skill for optimizing the product specifically for the hackathon.
These helped us improve the quality of the outputs during the Inception phase.
https://github.com/mashharuki/AWS-SummitHackathon-2026/tree/main/.claude/skills/hackathon-strategist
Environment Setup Matters!
The second thing we did was manually set up the development environment in advance.
Specifically, we prepared the pnpm monorepo project setup and implemented the minimum base structure for the frontend, backend, and CDK projects before running the AI-DLC implementation phase.
Environment setup places a very heavy burden on AI, so we handled that part ourselves on the human side and let the AI focus on logic design and implementation.
We also thoroughly prepared Agent Skills for the technology stack we planned to use and expanded our subagents before moving into implementation.
We kept the number of MCPs to a minimum and focused mainly on using AWS MCP Server, Serena MCP, and CodeGraph MCP.
https://github.com/mashharuki/AWS-SummitHackathon-2026/blob/main/.claude/.mcp.json
As a result, this preparation allowed us to steadily develop the product toward our goal, even under an extremely busy schedule.
Using Lean to Mathematically Ensure Program Safety
The final thing we did was incorporate a formal verification phase using Lean into AI-DLC in order to mathematically ensure the safety of the logic we planned to implement.
https://github.com/leanprover/lean4
Lean is a programming language developed for theorem proving, which is one of several approaches used to achieve formal verification.
I believe many people practice standard test-driven development together with spec-driven development, including AI-DLC.
We decided to go one step further and integrate formal verification checks into that process.
Unlike tests, formal verification proves the absence of bugs using mathematical formulas, allowing exhaustive verification. This is one of its major characteristics.
In fact, some source code auditing companies use formal verification in their work.
Our product was designed to handle privacy-sensitive information such as Slack chat histories, account names, and email addresses.
Because of that, we planned to implement encryption logic to protect this information.
However, there was a potential bug where the output values of the hash function used for encryption could match under certain conditions. We were able to discover this issue in advance through formal verification using Lean.
By incorporating a formal verification phase, I believe AI-DLC became an even more robust development methodology.
Of course, we also thoroughly researched formal verification and Lean in advance, and prepared subagents and Skills for them.
Conclusion
That concludes my summary of what we did to survive and succeed in the AI-DLC hackathon.
There are many different approaches for achieving goals with AI, such as cc-sdd, plan mode, and the /goal command. Among them, I feel that AI-DLC was proposed especially with practical enterprise use in mind.
After using it extensively in this hackathon, I would like to actively consider applying it in real-world development projects as well.
Final Thoughts
I am truly grateful to my teammates who fought alongside me, Katryo-san for mentoring us, and everyone from the hackathon organizing team.
I sincerely appreciate the opportunity to pitch on such a major stage as AWS Summit Japan and to stand on stage as a finalist.
I will continue taking on new challenges together with AI-DLC!
Top comments (0)