Building MyKMHub through AI-assisted development

Framing

More deliberate than vibe coding

“Vibe coding” is a recognizable shorthand for creating software through conversation with AI. It does not fully describe this project. MyKMHub uses written specifications, reusable content models, accessibility requirements, design-system constraints, automated tests, Git history, and explicit publication checks.

AI accelerates translation from intent to implementation. It does not own the product direction or determine whether the result is useful, accurate, accessible, or ready to publish.

Responsibilities

Human and AI responsibilities remain distinct

I define the purpose, content, information architecture, accessibility expectations, priorities, and acceptance decisions. Codex helps inspect the repository, propose and implement changes, run validation, maintain documentation, and operate the Git workflow.

Review is collaborative and iterative. When an implementation technically follows a request but misrepresents the intended visual or conceptual model, the implementation is corrected—not defended merely because it passes a test.

Workflow

A chain of independently verifiable states

1. Define or refine the requirement in project context.
2. Inspect the repository before changing it.
3. Implement against shared components and content models.
4. Review the visible experience and interaction.
5. Run lint, production build, and relevant browser tests.
6. Commit a coherent source milestone.
7. Push the commit to GitHub.
8. Verify Vercel and the changed production route.

Each transition matters. A local commit is not a GitHub update, and a GitHub update is not proof that production deployed successfully.

Guardrails

Accessibility is part of the implementation contract

Spectrum 2 supplies established components where viable. MyKMHub adds semantic page structure, keyboard access, visible focus, responsive behavior, appearance and reading preferences, and constrained theme controls with contrast checks.

Automated WCAG checks are a regression baseline, not a substitute for keyboard, screen-reader, zoom, reflow, contrast, and usability review. A passing scan cannot establish that a design decision is understandable or appropriate.

Knowledge management

The repository is also organizational memory

Product intent cannot live only in an AI conversation or editor session. The specification, roadmap, structured registries, operations guide, tool catalog, decisions, tests, and commit history preserve enough context to resume work with a different tool or on another device.

Content is maintained once and surfaced through relationships. Tools, portfolio evidence, practice notes, efforts, procedures, and environments can remain distinct records without becoming disconnected copies.

Value and limits

Where AI helps—and where it does not

AI is effective at repository analysis, implementation drafts, repetitive normalization, cross-linking, test creation, troubleshooting, and documenting operational knowledge while it is fresh.

It remains vulnerable to plausible misunderstandings, misplaced visual layers, overconfident assumptions, and technically valid solutions that miss the user’s mental model. Direct inspection, explicit evidence, and human correction remain essential.

Lessons so far

Treat the workflow as a governed system

• Give the repository and specification authority over chat memory.
• Record decisions and operational lessons as they occur.
• Separate preview state, saved state, and deployed state.
• Verify the exact layer or interaction the user described.
• Keep alternate tools viable through standard files and Git.
• Never confuse automation speed with product certainty.

This practice note will evolve as MyKMHub gains structured knowledge views, private administration, stronger backup and recovery, and measurable evidence about maintenance and publishing outcomes.