A Software Engineer Blogged Without Typing - ChatGPT Wrote 95% of It

A Software Engineer Blogged Without Typing

ChatGPT Wrote 95% of It.

Kevin Jalbert wanted to publish a blog post without sitting down to write one. So he picked up his iPhone and started talking.

What came out the other end, by his own accounting, was a post that was roughly 95 percent composed by ChatGPT — with the remaining 5 percent reserved for "minor adjustments or clarifications" he made himself. The piece appeared on his personal site under the title Using ChatGPT and Whisper: A New Approach to Blog Writing, and it's now a small but instructive case study in what mobile-first AI blogging actually looks like when a working engineer designs the workflow.

The tools, in order

Jalbert's stack has three pieces, and only one of them is famous.

The first is Whisperboard, an open-source iOS app that runs OpenAI's Whisper speech-recognition model locally on the iPhone. He uses a medium-sized Whisper model — large enough for good accuracy, small enough to process on-device without an internet round-trip.

The second is his own voice. He talks through the topic he wants to cover, recording into Whisperboard the way most people leave themselves a voice memo. There's no script. No outline he reads from. Just an extended verbal brain-dump.

The third is ChatGPT. He pastes the raw transcript into the model and asks it to reformat the dump into paragraphs.

That's the entire pipeline.

A deliberate technical choice: grounding the model

Jalbert, a software engineer by trade, was explicit about why the workflow is structured the way it is. The Whisper transcript isn't just a convenience. It's a grounding technique.

Large language models, he noted in the post, "can occasionally produce false information" — the polite term for hallucination. By providing the model with extensive source material in the form of his own transcribed voice, Jalbert reduces the surface area where the model has to invent. ChatGPT isn't asked to write a post about a topic. It's asked to clean up a post that Jalbert has already verbally written.

The result is content that reflects his actual thinking, formatted by the model rather than originated by it.

"By providing additional information and context to the LLM," he wrote, "I could guide the model's responses and reduce false information."

The 95 percent number, in context

The "95 percent composed by ChatGPT" figure has been the line that traveled. It deserves a second look.

What ChatGPT composed, in Jalbert's account, was the prose — the paragraph breaks, the transitions, the sentence-level polish that turns spoken thought into written text. What it didn't compose was the substance. The substance was Jalbert's transcribed monologue.

That distinction is the one his post quietly insists on. The model didn't produce a post. It produced a presentation of a post he'd already verbally written.

Local Whisper, local control

The choice of Whisperboard over a cloud-based transcription service matters more than it might first appear. Running Whisper on the iPhone keeps the audio on the device. There's no upload to a third-party server, no API key billing per minute, no dependency on network conditions.

For a working engineer with privacy instincts and an interest in keeping costs predictable, the local model is the obvious pick. It also makes the workflow function on a subway, a plane or anywhere else the cellular signal is unreliable.

The medium-sized Whisper model strikes a balance: the small models miss too many words for blog-grade output, while the largest ones bog down on a phone. Medium has become the working creator's default.

An honest assessment of what the workflow is good for

Jalbert's post stops short of claiming the technique works for every kind of writing. What he argued, in the closing of the piece, is that the approach is valuable for "quickly transforming a brain dump into text."

That phrase — brain dump into text — is the boundary of the workflow. Jalbert isn't using ChatGPT to research a topic, find sources or generate insight. He's using it to convert speech into prose. The thinking happens before the model gets involved. The model just formats.

That's the limit, and Jalbert names it explicitly. The technique works because he's already done the work that the model can't.

The implication for technical bloggers

The dev community has been cautious about AI-assisted writing for predictable reasons: hallucinated APIs, fake function signatures, code samples that compile in the model's imagination but nowhere else. Jalbert's workflow sidesteps all of it. The model is never asked to know anything. It's asked to format something the human already knows.

For other engineers contemplating a mobile blogging workflow, the recipe is unusually portable. A free open-source iOS app. A ChatGPT subscription most developers already pay for. An iPhone. A topic the writer can talk about for ten minutes without notes.

What it produces, on the other end, is a publishable post the writer didn't have to type.

A specific kind of leverage

Jalbert's piece doesn't argue that AI will replace technical writers. It argues for something quieter: that the keyboard is no longer the bottleneck, and that engineers who can talk through a topic now have a path to the published page that doesn't require them to retype it.

The 95 percent figure, in that light, isn't a triumph over the writer. It's a transfer of the labor the writer was doing inefficiently to a tool that does it instantly — leaving the writer free to do the part that only the writer can do.

Source: Kevin Jalbert, Using ChatGPT and Whisper: A New Approach to Blog Writing, kevinjalbert.com. https://kevinjalbert.com/using-chatgpt-and-whisper-a-new-approach-to-blog-writing/