How I Turn a Voice Memo Into a Restaurant Review Without Making AI Slop
The Claude Code skill chain behind All Around STL: a voice memo and phone photos in a Google Drive folder become a finished restaurant review, with every sentence traced back to something I said or photographed.
These days a restaurant review on All Around STL starts the same way every time. I eat somewhere, I talk into my phone about it, and I take a bunch of pictures. Then Claude Code turns that into an article.
That sounds like a recipe for slop. An AI can write a convincing review of a restaurant it has never been to, and the internet is full of them.
Ahrefs recently wrote up how they use AI on every article without making slop. Their test stuck with me: slop is what you get when nobody supplied the raw material and nobody owns the result. This is my small-scale version of that, using one real review as the example.
The rule: the AI never tastes the ramen
I went to Kishimoto Mendo, a ramen bar in the Loop, with a buddy before grabbing beers next door. Everything in the published review traces back to two things: what I said into my phone afterward and what I photographed. If a detail isn’t in one of those, it doesn’t go in the review.
That one rule shapes the whole pipeline. It’s a chain of Claude Code skills, each doing one job and handing a file to the next, with a few places where it stops and waits for me.
1. Start with raw material only I have
The input is a Google Drive folder per restaurant. I drop in the voice memo and every photo off my phone, unnamed and unsorted.
I talk instead of type on purpose. When I type a review I clean it up as I go and lose the stuff that makes it mine. Talking keeps the hedges and the tangents, like rating it “low eight, eight-three, eight-four” because I couldn’t decide.
The photos are evidence too. The name-and-upload-photos skill looks at each image and renames IMG_xxxx.jpg to something like counter.jpg or receipt.jpg. Drive renames are visible to anyone with access, so it shows me the full rename list and waits before touching anything.
2. Transcribe it locally and fix the names by hand
The transcribe-review skill runs OpenAI’s Whisper (the small.en model) on my laptop. No API key, nothing leaves the machine, and it writes a plain transcript.txt.
Whisper is good, but it doesn’t know St. Louis. It heard “Menya Rui” as “Menure” and “Nudo House” as “New Doe House.” Those got fixed by hand.
That’s the only editing the transcript gets. The skill’s rule is to fix obvious mishearings and never add facts, because every later step treats this file as ground truth.
3. Plan before anything gets drafted
Drafting is the expensive step, so the plan-review skill reads the transcript first and writes a plan.json. It holds the rating, the intro hook, the beats for each section, a list of proper nouns the drafters are allowed to use, and a list of gaps.
The gaps list is where it earns its keep. In the memo I said the ramen had “four layers of depth of flavor” and then only named three. The planner caught that and wrote “don’t invent a fourth.”
It also asked questions instead of guessing. Did I get there closer to 6 or 7? What did my buddy order? The review ended up saying “early evening” and only talks about my bowl.
This is the first real stopping point. I read the plan and confirm the rating before any drafting starts, because re-planning is cheap and re-drafting isn’t.
4. Draft against my own writing, not the internet’s
Four drafters run in parallel, one each for the intro, experience, food, and atmosphere sections. Each one gets the plan plus a bundle of examples pulled from about 100 reviews I wrote by hand in 2024 and earlier.
write-review skill. Rule 8 is the one that keeps it sounding like me.The bundle also carries a list of banned phrases (“culinary journey,” “nestled in,” “must-try”) and a voice guide built from my old reviews. The point is to sound like me, not like a nicer average writer.
The skill says that directly: lift my words from the transcript where they fit and don’t upgrade my casual phrasing into smoother prose. Fragments and slightly-off grammar stay if that’s how I said it.
5. Score it, critique it, and let it fail
Every section then goes through score.py, a deterministic checker. It looks for banned phrases, missing HTML, length, how far the rhythm drifts from my old reviews, and grounding, meaning whether every proper noun appears in the transcript.
The food section scored 0.0 on grounding because “Tonkotsu Black Ramen” isn’t in the memo. In the memo I literally said “whatever the name of it is that’s on the receipt.” The name and the $17 price came off the receipt photo.
That’s a known false positive, and I’ll take it. A checker that sometimes flags a real fact is better than one that lets an invented dish slide.
Normally a second layer kicks in. The critique-review skill spawns one critic per section to catch prose that passes the numbers but reads flat, then a reviser gets up to three tries. If a section still fails after three, it stops and hands it to me instead of looping forever.
6. Humanize last, and keep the bad grammar
Next, a polish pass reads all four sections together for repeated phrases and tone drift. Then my humanize-prose skill strips AI tells like “not just X but Y” constructions and em-dash overload.
The humanize step runs under the same voice rule as everything else. It only changes a line if the line reads like AI, not because a more correct version exists.
The drafters lift my lines straight from the transcript, and nothing downstream is allowed to smooth them out. That’s how “I love me some ramen” made it to the live page, along with “100% I’m going back here.”
7. Photos go where they belong, and the receipt goes last
The add-images skill looks at every photo (it opens the actual JPGs rather than trusting filenames) and places each one next to the paragraph it shows. It also picks the feature image. When the food carried the visit, that’s the hero dish, which is why the ramen bowl leads this one.
Every review ends with a photo of the receipt. I always bring the receipts, literally.
Then an internal-linking step connects mentions like Blueberry Hill to my existing reviews, and push-to-drive creates a Google Doc in the same Drive folder the photos came from.
8. I read every word before it goes live
The Google Doc is the handoff. I read it top to bottom, and nothing publishes until I run publish-new-article with an explicit --write flag.
My name and my face are on that page, so I’m the one who has to stand behind the 8.3.
Is it working?
The only thing I’ve done on All Around STL this year is run this pipeline. The first review it wrote went live in September, and I’ve run it on 11 reviews so far.
The site fell off a cliff in December 2025 and sat there for most of this year, at a handful of clicks a day from Google. In early November it came back, peaking near 190 clicks in a day and holding above 100 since.
I can’t tell you exactly which pages Google decided it liked. But nothing else on the site changed, and this is the first real movement since last December.
What broke on this one
Every run appends a line to a learnings log, including the parts that went sideways. This is the real entry for Kishimoto Mendo.
The planner was too cautious. It marked the food and atmosphere sections as “needs more,” so the drafters wrote placeholder stubs, and those two short sections got written straight from the transcript instead.
The drafters also stuck in their own image tags, including one with the wrong alt text, even though they’re told not to. Those got stripped before the image step ran.
Because the draft was so short and scored zero AI tells, this run skipped the critic, polish, and humanize agents entirely. And the photo upload to my host failed, so the images went up separately.
None of that made it to the page, because the gates caught it. But the pipeline isn’t hands-off, and I’d be lying if I said it was.
The point
The Ahrefs piece argues the effort moves upstream when drafting gets cheap. In this pipeline the parts I can’t hand off are all upstream: recording a useful memo, taking the right photos, and reading the plan.
There are 105 voice-memo folders sitting in my repo, and every one is a place I actually went.
The bottom line
Claude Code does the typing. I eat the ramen and make the call on the rating, and nothing goes live until I've read it.