Recipe · · 4 min
How Claude co-wrote 71% of Shhots AI's site commits and kept out the slop
Claude writes most of the Shhots AI website now, from product pages to price comparisons. Written rules, automatic checks and a person on every merge keep it from reading like AI.

- Company
- Shhots AI, AI ad generator
- Problem
- Hundreds of pages to write and keep current, and buyers who can spot AI writing
- Install
- Written rules Claude reads first, checks on every draft, a person on every merge
- Result
- 71% of site commits co-written by Claude in Aug 2026
The problem
Shhots AI sells to marketers, and marketers have read more AI-written copy than anyone. A site that reads like AI loses them in a sentence. Google demotes it too: its updates go after templated pages with undated prices and nothing first-hand.
The site still needs a lot of pages: one per ad platform, per AI model and per use case, plus comparisons, pricing posts and launch posts. Each one has prices and facts that go stale. With one founder, that work waited.
What we built
Claude writes the change, the checks run, and I review the pull request before it merges.
- Rules Claude reads before it writes. A project guide in the repository tells Claude how the site is built and what every page must carry. Copy rules set the voice and list banned phrases. Brand and front-end rules cover design. An SEO playbook says what can change on a page and when.
- Checks on every draft. A lint flags banned words, em dashes and padding. The rules require descriptive alt text on every image. Every number on the site comes from one file that records where it’s shown and what backs it, so a page can’t make up its own stats.
- Facts checked at the source, with a date. Claude checks each competitor price on the competitor’s own pricing page and writes down the date. If the page can’t be read, the post waits until a person has read it in a browser. Every “best tools” list that includes Shhots says we make it.
- AI audits of the whole site. A script has a small AI model compare every pair of pages, looking for two pages that answer the same search. One run covered 114 pages in 2,229 pairs. It cost 13 cents and found 34 pairs for a person to review.
- A person on every merge. Every change lands as a pull request, and a person reads it before it goes live. There have been 188 so far.
How the rules are written, and what the checks look for, is the part we keep for clients.
What it cost
- Tools: a Claude subscription, plus GitHub and Astro, which are free at this size. Claude’s cost is not included here.
- People: one founder reviewing pull requests.
- Not included: the product itself, video production and paid ads.
What came out
| Before | After | |
|---|---|---|
| Site commits co-written by Claude | 0 of 63 (Feb to Apr 2026) | 70 of 98, 71% (Aug 2026) |
| Since Claude started | Not applicable | 242 of 466, 52% (May to Oct 2026) |
| Where the site’s numbers live | Typed into each page | One file every page reads from |
| “Best tools” lists that say we make Shhots | 1 | 40 (24 Aug 2026) |
| Pages checked for overlap with each other | Never | 114 pages for 13 cents (22 Sep 2026) |
These numbers come from the Shhots AI website’s git history. We counted a commit only when Claude is tagged as a co-author.
The test of quality is what readers and search engines do with the pages. After the September fixes, Google clicks had their best days since early July on the same number of impressions, and ChatGPT brought nearly as many new users as Google search. That story is in how Shhots AI grew its new users from ChatGPT by 684%.
What we learned
Write the rules first, then go fast. In July, Claude helped ship quickly: six new posts in one commit, and two posts merged into one. What it needed next was a written list of what every page must carry.
In August we wrote that list. A review of the site’s 39 “best tools” lists set the bar: dated prices, first-hand testing and a clear owner for every search. We folded 9 lists into stronger pages, rebuilt most of the rest with dated prices, and restored the merged roundup at its old address.
Claude now reads those rules before every change, and one page gets one change at a time, measured for 14 days before the next.
Installed on your company
This fits any firm whose website, content or product pages wait on one busy person, or that tried AI writing and got copy nobody wanted to publish. It works the same way as Mockey’s setup.
In week one, we write your rules with you: voice, banned phrases, what every page must carry, and where your numbers come from. Then we connect Claude to your site’s repository, set up the checks, and put your reviewer on every merge.
Each month you see pages shipped, the share written by AI, what the checks caught, hours spent reviewing and the tool bill. It is part of AI for product and AI for growth. If your site waits on one busy person, pick a time in the calendar below.
Rajat, 4 Nov 2026. Running since May 2026. Tools: Claude Code, GitHub, Astro, Search Console, GA4.
This work sits under AI for product.


