elizatrinity.tech launch series · Part 3
It Used to Take Me Hours to Publish One Article. In 18 Days, I Had 49 Live.
Part 2 ended with the pipeline running end to end — download, transcribe, structure, draft, review, publish, with me only touching the one deliberate gate: approval.
This part is where I put a number on what that actually changed.
The Old Math
I described the manual era in Part 1: copy ChatGPT output, open Wix, paste and reformat, go back for a header image idea, generate it separately, ask for alt text, ask for SEO keywords, back to Wix, assemble everything by hand, publish.
I went back through my own ChatGPT history to time it properly, on one real example — the article "Are the Blessed Cursed?":
- 1:55 pm — I gave ChatGPT the transcript.
- 2:02 pm — First output back: title, main content, prayers, golden nuggets, visual asset suggestions, and alt text. Seven minutes.
- 2:02 – 2:35 pm — Generating each image one at a time (redoing any that came out wrong), then alt text, then an excerpt — which meant switching over to Wix first just to screenshot something to work from. Thirty-three minutes.
That's 40 minutes, measured, and it only covers the ChatGPT side. After that came a second phase I didn't time as precisely: paste everything into Notion for my own records, then copy it all into Wix — paste the YouTube URL manually, check that every heading came through as H1/H2/H3 correctly, upload each image to the right spot one at a time, paste the alt text, paste the excerpt.
Based on what that phase actually involved, my rough estimate is another 2 to 3 hours for that assembly step. It wasn't just pasting — every paste had to be checked against the Wix editor's own layout rules: alignment, borders, image frames, spacing that didn't carry over cleanly and had to be fixed by hand, one block at a time. That puts the whole thing at roughly 3 to 3.5 hours per article, start to finish.
Same sermon, two eras. Here's the old Wix blog post for "Are the Blessed Cursed?":
And the same teaching today, on revealedinsight.org:
The New Math
Here's what's actually in the repo, not a rounded marketing number.
The aios-workspace repo — the one holding the whole pipeline — was initialized on March 1, 2026. Eleven days later, on March 12, the website went live end to end: every page, every component, and the first sermon ("Mammon") published, pipeline stages 0 through 9 operational.
A week after that, on March 19 — eighteen days after that first commit — the internal log for that session reads "Total 49 Live." Forty-nine study notes, published, in eighteen days, working weekends only around a full-time program.
Today, as of this writing: 379 published study notes.
And per sermon, the pipeline itself — stages 0 through 10, unattended, not counting how long a draft sits waiting for me to open the review dashboard — runs in roughly 25 to 40 minutes for a typical 1.5–2 hour sermon. That works out to about 0.3× the sermon's own length: for every hour of audio, about 17–18 minutes of processing.
| Sermon (audio length) | Words transcribed | Wall-clock time |
|---|---|---|
| The Danger of Covenants (92 min) | 11,217 | 25 min 40 s |
| The Untold Secrets of Communion (115 min) | 14,660 | 32 min 11 s |
| The Danger of Covenants Pt. 2 (129 min) | 17,935 | 38 min 16 s |
Where that time actually goes, from one of those runs: about half of it — 15 minutes — is Stage 0, downloading and transcribing with Whisper's smallest model on ordinary laptop hardware, chunked so it doesn't run out of memory. Cleaning and segmenting (Stages 1–3) take about 6 minutes combined. Structuring the sermon's theology (Stage 5) takes about 4. Writing the actual draft (Stage 7) takes about 6. Everything else — topics, quotes, knowledge graph, image prep (Stages 6, 8–10) — finishes in under 2 minutes combined.
(Worth saying plainly: this is three runs on one laptop, timed from log and file timestamps rather than built-in instrumentation, so treat it as typical, not a lab benchmark. It also only covers the unattended processing — it leaves out the review, the approval, and any image I generate by hand when the automated match isn't good enough.)
What Actually Changed
It's not that the new version is faster at any single step — transcription on a memory-constrained laptop is still, on its own, one of the slower parts of the pipeline. It's that none of the ten manual handoffs exist anymore. Nobody is copying anything between ChatGPT and Wix. Nobody is asking for a header image idea and then switching tools to generate it, then switching again to paste it in the right spot. The pipeline does the parts that used to be typing and clicking, and I do the one part that's actually a judgment call: deciding whether a draft is good enough to go live.
That's the real difference between 3-plus hours for one article and 49 live in 18 days. Not a faster me. A version of the workflow where most of it isn't me anymore.
(Article 4 — How I Work With Claude Code — linked here once it goes live.)