On this page
And in just one day
I spent 15 years on the bridge of ships. DP operations, offshore support, complex systems where documentation is operational truth rather than paperwork. Manuals, capability analyses, compliance records. I understand this documentation well. I understand how it should work, what it needs to do, and where existing tools fall short.
What I do not have is 15 years of software engineering experience.
I am learning. I have built small things. But I have never shipped a full-stack application from scratch. Until today.
Manifest is a publishing platform I designed for maritime technical documentation. The core idea is that content exists as reusable blocks (sections, data tables, diagrams), and documents are assemblies that pull those blocks together. Write a section once, use it in multiple publications. Update a table, and every document referencing it reflects the change. Pin a version when you publish, so the record stays stable.
Today I took Manifest from a specification document and an empty repository to a working system. Database, API, admin interface, public reader, deployment configuration. 113 files. Roughly 9,400 lines of code. In one day.
The process
I did not write most of that code by hand. The workflow was a collaboration between two AI systems and me.
Claude wrote the software requirements document: a comprehensive specification covering data models, API contracts, edge cases, and architectural decisions. From that document we derived a task breakdown and progress tracker that would guide implementation.
I then had Codex review Claude's assumptions. Where had Claude been too optimistic? What edge cases were missing? What would break under real usage? We revised the plan based on that critique.
Once the specification was final, Codex implemented it. I described what needed to happen based on the spec, and Codex generated Python, TypeScript, and configuration files. I reviewed the output, identified issues, and directed corrections. The implementation moved quickly.
At the end of the day, Claude performed a security audit of the codebase. That audit produced a list of issues that need addressing before Manifest can go to production.
I contributed the domain knowledge throughout. I understand how maritime documentation actually works. I understand what problems needed solving. I made the decisions and identified the moments when the AI output did not match what I actually needed.
Observations
The AI did not replace the need to understand what I was building. It replaced the years of implementation experience I do not have: the ability to quickly scaffold a FastAPI application, wire up a React frontend, configure Docker and Nginx. Tasks that would have taken me weeks to learn through trial and error.
I still had to understand the concepts. I still had to make judgment calls. But I was not bottlenecked by syntax or framework-specific knowledge.
Limitations
This would not have worked if I did not know what I wanted to build. The AI can generate code, but it cannot tell you whether your architecture makes sense for your use case. It cannot tell you whether your data model will create problems later. It cannot tell you what good looks like in your specific domain.
If I did not have domain expertise, if I did not understand maritime documentation requirements, I would have built something that looked like software but did not actually solve the problem.
The AI amplified what I already knew. It did not replace the need to know things.
Implications
Two years ago, building something like this would have required either months of self-teaching or hiring someone with the skills I lacked. The barrier was not ideas. It was implementation capacity.
That barrier is now considerably lower.
I am uncertain what this means for software engineering as a profession. But I know what it means for me: I can build tools I have been conceptualising for years, without waiting until I have accumulated sufficient implementation experience.
Current status
Manifest currently runs on my local machine. It has bugs. It has security issues that need addressing before it can be deployed publicly. Moving it to a production server will require careful work.
But the foundation exists. I can navigate the admin interface, create content blocks, assemble them into documents, and see the result render in the reader.
It is not complete. But it is functional.