From 3,370 old posts to a versioned writing workflow in three weeks
A case study in turning a long-running personal blog into something an AI collaborator can work with, without losing the author's voice
At a glance
| Situation | A personal blog running since 2007, with about 3,370 posts, a WordPress export as the only structured copy, and a second site the author wanted to write on seriously |
| Approach | Extract the archive to Markdown, write a style guide from the author's own posts, give each article a git history, and keep Claude as a drafting and research partner under written rules |
| Window | 4 to 21 October 2026, with commits on 13 of those 18 days |
| Output | A 3,381-file Markdown archive, a style guide, a post template, 7 published pieces on two sites, 4 drafts in progress, and 93 commits recording how each piece evolved |
The situation
The blog, On The 8 Spot, started in 2007. By 2026 it held thousands of posts in one WordPress database, and the only portable copy was a 14 MB export file. A second site, a collaborative essay blog, was where the author wanted to write policy pieces in earnest.
Three problems stood in the way of writing well with an AI assistant:
- Nobody, including the author, knew what was in the archive. Posts ranged from long personal essays to quoted articles and video links, all in one pile.
- Claude had no sense of the author's voice. Left alone it would write competent, generic prose.
- There was no record of how a piece changed. WordPress keeps revisions, but they don't show why something was cut or which claim was checked.
How the work was run
- Read the export, don't load it. The export is too large to read whole, so it was parsed with a script and the posts were written out one per file, named by date and slug, with title, date, categories, tags and the original URL in the front matter.
- Sort by who wrote it. Using the blog's own categories, the 3,381 files were split into the author's own posts (1,124), reposted articles by other people (1,445), everything else (803) and the 9 pages. That split matters later: only the first group says anything about the author's voice, and the reposts must never be mistaken for it.
- Write the style guide from the evidence. A read of the author's own posts, plus guest posts on the second site, produced a guide covering the core voice, signature moves, habits to tighten, post types with rough lengths, and instructions for Claude. It states its own limits: most of the strongest first-person writing is from 2008 to 2011.
- Give each site a folder and each post a lifecycle. Each site has
drafts/,published/andassets/. One Markdown file per post carries a status, a publish date and the WordPress post ID once it is live. - Commit at every stage. One post per commit, with messages like
research:,draft:,revise:andpublish:, so the history reads as the article's biography. A private remote keeps a backup. - Write the rules down where the assistant reads them. A short project file tells Claude which site it is writing for, to read the style guide first, to suggest edits rather than rewrite, and never to publish without confirmation.
What the history shows
Of the 93 commits, the verb in the message gives a rough picture of the work:
| Kind of commit | Count |
|---|---|
revise: |
44 |
research: |
17 |
draft: |
11 |
publish: |
8 |
docs: |
3 |
| Other (setup, tooling) | 10 |
Revision outnumbered first drafts four to one. Many of the revisions are fact checks, such as confirming a figure against the full text of a government order, replacing a claim that could not be supported with a safer one, or adding a source. Others record the author's own corrections, such as a note that an unnamed friend's story should stay anonymous, or a replacement ending in the author's words.
The output so far is 11 posts and drafts totalling about 23,000 words, across the two sites. Seven are published and four are in draft.
Where it went wrong, and how it was caught
- Scaffolding left in the post. The style guide records an older post that still opens with a pasted line from an AI tool. The rule that followed: never leave tool chatter in a draft, and run a final proofread pass.
- Claims that sounded right but weren't backed. A draft's opening claims were flagged as unsupported before publication, and several commits replace a confident sentence with a sourced or softened one.
- The author's voice versus the assistant's. The guide tells Claude to flag the author's known habits (run-ons, promises of future posts, a buried thesis) gently, and to keep the habits that make the writing his, such as a short honest opening and a tight "TLDR of my proposal" list.
- Personal data in the export. The WordPress file holds author and commenter emails and IP addresses. The archive was built without comments or author data, the export itself is excluded from git, and a private analysis of the archive is kept local and never copied into public drafts. One phone number inside a reposted article was redacted.
What was left open
- The automatic HTML-to-Markdown conversion is imperfect for tables and embeds, and images still point at the original WordPress URLs.
- The 803 files under "everything else" are not reviewed. Not all of them are other people's writing.
- Publishing to the personal site is still manual, because its tools are not connected to the assistant.
- Four drafts are not yet published.
Takeaways
- Know what you have before you write with it. Sorting the archive by authorship made the style guide possible and kept other people's words out of it.
- Put the voice in writing. A style guide drawn from the author's real posts, including its own limits, gave the assistant something to follow other than its defaults.
- Commit the thinking, not just the result. Stage-by-stage commits turned the history into a record of what was researched, checked and cut.
- Make the assistant suggest, not decide. The author keeps the final say, and nothing goes live without confirmation.
- Treat personal data as a first-class risk. Excluding the export, stripping comments and keeping private analysis local were set up on day one, not added after.
About these numbers
File counts come from the archive's own README and a directory count. Commit counts and dates come from the repository history. Word counts are from the post files and include drafts. This case study does not report session time or message counts, because the transcripts were not analysed.