How it works
Contengi runs a small team of models, each on the job it does best, with your knowledge base in the middle and four filters between the first draft and you. This is the whole flow.
Every model in Contengi reads from the same brand file: your strategy, point of view, audience, tone of voice, writing guide, hard rules, approved sources, and sitemap. We build it with you during setup, and it's the reason a quick LinkedIn post and a 2,000-word guide sound like the same brand.
Each model gets the slice it needs and nothing more. The writer gets your voice and rules, research gets your competitors and approved domains, and the link step gets your sitemap. Less to read means less to pay for, and less for a model to get distracted by.
One model doing everything is either expensive or sloppy, so Contengi splits the work. A fast model does the looking up, Claude makes every call that needs taste, a checking model reads what comes out, and plain code handles the rules with one right answer.
Fast and cheap, so it takes the legwork. It never writes a word your readers see.
Anything that needs judgement runs here, with your full brand voice loaded.
A model built to check writing, so it gives a straight yes or no for each pattern.
Where a rule is black and white, a model has no business deciding it.
What we learned the hard way
We tried giving blog editing to the cheaper model too. On a real post it added a TL;DR the article already had and put new content above the title, so editing went straight back to Claude. We tested moving workflow research over as well, and kept it on Claude because checking facts is judgement. The fast model earns its place in conversation and lookups, where it's quick, cheap, and more than good enough.
AI writing gives itself away in patterns, and a banned-word list only catches the obvious ones. So every piece goes through four filters, each one catching what the last one can't.
1In the brief
Every writer starts with Contengi's anti-slop standard and your brand's hard rules in its instructions. They're on for every brand and can't be switched off.
2The editor's pass
Claude re-reads the draft as an editor, strips the patterns it finds, and keeps every fact, example, and link where it was. The length has to stay within 10% of the draft.
3The exact scan
Plain code counts every em dash, emoji, and banned phrase, and swaps em dashes for spaced hyphens. No model gets a vote on a rule that is black and white.
4The slop check
Jev reads the finished piece for the tells a word list misses, like a sentence that sets up a straw man just to knock it down. It points at the exact sentences, Claude rewrites those, and Jev checks again.
A real fix from a real run. Only the flagged sentence changes, so the parts that were already good stay yours.
Your Claude key pays for the writing, at Anthropic's own prices with nothing added on top. Everything marked included runs on Contengi's bill.
Every provider we use is listed in our privacy policy.
FAQ
Three, each on the job it does best. OpenAI's GPT-6 Luna runs the Marketing Assistant's conversation and lookups, Anthropic's Claude does every piece of writing, editing, and research that needs judgement, and TypeSafe's Jev checks finished text for AI slop. Plain code handles the rules that are black and white, like em dashes and dead links.
Yes, for the writing. Drafts, rewrites, and edits run on your Anthropic key, so you always get Claude at full strength on the words your readers see. The assistant's conversation, lookups, and the slop check run on Contengi's own AI, included in your plan.
A model built to check writing. Contengi asks it a set of yes-or-no questions about a finished piece, one per known AI pattern plus your brand's never-rules, and it points at the sentences that fail. Only those sentences get rewritten.
No. It only touches the sentences it flags, and a fix has to match one exact spot in the text or it is skipped. Everything else stays exactly as written, links and formatting included.
Nothing stops. Your Usage page shows how much of the month you have used, and if you ever use it all, the assistant carries on through your own Claude key until the month resets.
No. Your writing runs on your own Anthropic key, which Anthropic does not train on, and the included AI runs with zero data retention. The full list of providers is in our privacy policy.
Keep exploring
In the platform