← Back to Work

Evolving

Quiet Growing Archive

How I turned conversations with ChatGPT into a design intent, a bilingual archive, and a repeatable human-led review process.

Personal publishing systemSite record

Overview

This project began before the first page of code. I wanted a place that could keep my work, thinking, and growth together without starting over whenever my interests changed. I used conversations with ChatGPT to make that intention explicit, then asked a coding agent to turn it into something I could open, judge, and improve. My contribution was not choosing every CSS value: it was deciding what the site was for, what could wait, and when a version was ready to publish. The eight decisions below trace that human workflow, from questions to a working public archive.

Outcome / current state

The bilingual archive went live on September 24, followed by a first visual refinement on September 25. On September 28, I manually published the author profile, illustrated portrait, refined tools, and chapter-based project story to Cloudflare Pages. The live domain was checked for matching assets, image loading, language switching, and filters. The collection still distinguishes actual site records from illustrative material. Further portrait framing and About refinements are being reviewed locally, not yet published.

Decisions

Ask what should survive

I asked ChatGPT to collect my intent through short rounds of questions, rather than jump straight to a layout. Who was this for? What should a visitor take away? What would still matter in three years? My answers gave self and public readers a 50/50 weight, put practical experience before daily thinking and technical learning, and made career value a benefit rather than the main purpose. The useful output was a set of priorities, not a polished homepage.

Turn answers into intent.md

ChatGPT maintained a document named intent.md inside the conversation and revised it as I answered. It separated the long-term purpose, content priorities, boundaries, and open questions. English-first with a Chinese version was an explicit choice; so was preserving how a project evolved, not only its result. This was a conversational design document, not a repository file that generated the site. Its job was to give later requirements and design decisions something to be checked against.

Research parts, not a site to copy

I then asked ChatGPT to research blogs with a similar visual language and underlying idea. Its report proposed different references for different questions: Paco and Steph for restraint, Simon and Odhy for a time stream, Maggie for evolving notes, and Andy for connections. These were research suggestions, not a claim that I personally audited every site. The useful distinction was between events, understanding, and outcomes; the design could combine those lessons without copying an entire interface.

Give each kind of content a job

The architecture became Home, Work, Notes, Timeline, and About. Work explains outcomes and evolution; Notes carry analysis and developing judgments; a small fragment can enter Timeline without first becoming an essay. I wanted Work/Life and date-based views, not a taxonomy I would have to maintain constantly. The visual direction, Quiet Growing Archive, translated that intent into warm paper, graphite text, moss green, and restrained orange: readable and quiet, with time and change visible rather than a wall of portfolio cards.

Separate the long-term system from the first release

My immediate request was deliberately small: run a demo and test hosting and the purchased domain. I specified a blog subdomain, leaving the root domain alone. A Vite demo went to Cloudflare Pages and Spaceship DNS connected blog.ahatole.com. After that path worked, I returned to the earlier intent and requested a fuller bilingual version. I allowed clearly labelled sample material because I did not yet have time to supply all the real content. That kept the template testable without inventing a personal history.

Delegate execution with a concrete brief

I asked the lead agent to make planning and SPEC decisions from the earlier context, and to ask me when essential information was missing. The local SPEC turned intent into routes, content relationships, language behavior, and acceptance checks. Terra handled implementation; the lead agent coordinated and verified it. Shared bilingual data and static prerendering kept the public site simple to host and maintain. I set the purpose and constraints, while the agents worked out the implementation within them.

Review a page, then authorize a release

I asked to see the site locally before publishing and later requested Astra design reviews. Review suggestions became bounded iterations: a shorter homepage introduction, clearer heading hierarchy, month chapters, and quieter related reading. The lead agent checked both languages, routes, filters, history, and desktop/375px layouts; I gave the release instruction. Publishing to Cloudflare was followed by checking the actual domain. A successful local build, an uploaded folder, and a live release were treated as different states.

Keep unfinished work visible

When I had no time to develop the homepage narrative and author voice, I deferred them and prioritized the tool area. I also asked for this case study, then corrected its first draft because it described the build more clearly than the human process. AI Capture, automatic linking and synthesis, and comments are not built; automated translation updates are not implemented either. They remain intentions, not current features. For me, this workflow means answering questions, preserving decisions, judging concrete pages, and choosing the next scope—not treating one prompt as a finished product.

Evolution

  1. Vite demo published on Cloudflare Pages; blog.ahatole.com connected through Spaceship DNS
  2. Bilingual static archive published after tests and local browser review
  3. First visual refinement published and checked on the live domain
  4. Second refinement and lighter Notes and Timeline tools passed local review but were not yet live at this point
  5. Manually published the author profile and illustrated portrait; live pages and interactions verified