lately

I built a way to blog without typing, then stopped blogging

Brian Stever

2024 · Next.js, OpenAI, PostgreSQL, TipTap

The short version. I wanted posting to my own site to take one short voice memo and a button press. So I built a pipeline: Whisper transcribes the memo, GPT-4 turns it into a draft, DALL-E makes a header image, and a Next.js site backed by PostgreSQL publishes it. Technically, it worked. It didn't solve the harder problem, which is that I still had to want to blog.

1The blank page

I like having a personal site, and I like the idea of writing on it regularly. What I don't enjoy is the moment you open a blank document, put your fingers on the keyboard and have to produce coherent prose from nothing. Talking is easier. Talking I can do for free.

So the idea was embarrassingly simple. If typing was the bottleneck, take typing out. Record a voice memo, transcribe it, let GPT-4 turn the ramble into something publishable, generate a header image and ship the whole thing to the blog. In theory, publishing goes from “write a post” to “talk for two minutes.”

There are obvious reasons this is a bad idea. Your spoken thoughts are usually worse than you remember. And once you let an LLM “help,” it will now and then decide you're an authority on things you've never studied. Both still seemed easier to manage than a blank Google Doc.

2How it works

You upload an audio file and Whisper transcribes it. GPT-4 gets the transcript with a prompt asking for a coherent first draft, and DALL-E makes a header image from the title GPT-4 comes up with. The result is saved to a PostgreSQL-backed content store and shows up in a Next.js front end, where a TipTap editor lets me clean it up before publishing.

I kept the interface narrow on purpose. I wanted it to feel less like a CMS and more like a small machine: record, wait, review, publish. Every bit of configuration I added took me closer to rebuilding the writing tool I was trying to remove.

ai-blogger pipeline

  1. 01Record a rough voice memo
  2. 02Transcribe with Whisper
  3. 03Expand into a draft with GPT-4
  4. 04Generate a header image with DALL-E
  5. 05Persist title, body, and image URL
  6. 06Publish to the site
const createPost = async (audioBlob) => {
  const transcript = await whisper.transcribe(audioBlob);
  const draft = await gpt4.complete({ transcript });
  const image = await dalle.generate(draft.title);
  await db.posts.insert({ ...draft, imageUrl: image.url });
};
Figure 1. The pipeline, simplified. Press Run and it steps through transcription, drafting, image generation and saving the post.

3What happened

It worked well enough to prove the idea. A rough memo became a structured draft in under a minute. The posts were often readable and occasionally decent, and sometimes suspiciously eager to improve on what I'd actually said. OpenAI would call that “creative assistance.” I would call it inventing facts about my life.

The surprise was that taking out the typing didn't give me a blogging habit. The thing stopping me wasn't keystrokes. Sometimes it was taste, and sometimes energy. Sometimes it was not wanting to read a machine's version of my own half-formed thought and decide whether it still sounded like me.

Table 1. The prototype, in short.

MetricValue
External AI services3
Time from memo to draft<60 s
Rich-text editorTipTap
Days before I forgot to use it90

4Where it broke

First, transcription. Background noise, mumbling and low-energy voice notes got me Whisper transcripts that were technically English but not what I'd said. At one point “AI blogging” came out as “eye flogging,” which felt less like a typo and more like a judgment.

Second, the model drifting from what I said. GPT-4 is very willing to fill the gaps in your thinking with connective tissue you never gave it. That's great if you want a confident blog post, and less great if you need the confident blog post to stay attached to reality.

Third, maintenance. The project ran on a free-tier database, and I went long enough without publishing anything that the database paused itself for inactivity. It's an almost perfect case study in solving the wrong problem very well.

5The part it can't do

I still think the core idea is sound. For someone who thinks clearly out loud and is comfortable editing a generated draft, it's a genuinely useful workflow. It cuts the overhead of getting from a thought to a draft, and that matters.

What it didn't do was take the editor out of the process. The machine can draft, summarize and decorate, but it can't decide whether a post is worth publishing. That's still a human problem. Unfortunately, the human was me.