How we write our posts
We build a framework for AI agents, and we use AI to write about it, this article included. In short: a coding agent drafts from our docs and our code, lists each claim with its source and what it left out, and nothing goes out until we’ve edited it and said yes. Here’s how, and what we’ve learned doing it, in case it helps you do the same.
Draft from sources, not from memory
Our drafts come from a coding agent, because it can read what a claim rests on: our docs and our code. A model writing from memory sounds just as sure when it’s wrong. A coding agent can open the page or the file and check.
Ask for the claims, not only the text
Every draft comes with its list of claims, and where each one was checked: a docs page, or a file in the code. A claim with no source doesn’t stay: either the agent finds one, or the sentence goes.
For example: a title we drafted promised automatic rollback. Our docs say a rollback is one command you run, so “automatic” went.
Ask what it left out
Each draft also lists what we chose not to say: a feature that isn’t shipped, a number we can’t back. That list catches the confident sentence that would cost more trust than it wins.
Write down what never goes out
Some things stay out of every draft, whatever it’s about: a customer’s name, or anything told to us in private. Examples use a made-up company, Acme. The agent checks each draft against that list, and so do we before we say yes.
Let your edits teach
We edit every draft, often down to half, and what we change becomes a rule for the next one. Some of ours include: answer the question that was actually asked, and no exclamation marks, because we add those ourselves, where they belong. Each draft starts closer to how we talk.
Keep the yes with a person
Nothing goes out without our yes. As we put it in a reply under one of our first posts:
It’s public, it has my name on it, and whether it reads right is a judgment no rule could capture. An agent can draft it, but it goes out only after my edits and my yes.
The same shape, for your agents
If an agent of yours writes to customers, it’s the same shape: it works from your systems, not from memory, you choose which of its steps wait for a person’s yes, and every step is on record. We built Kindgi around that shape. It’s an agent framework for production: your coding agent writes the tools, agents and flows in your own code, and Kindgi runs them and keeps every step on record. More in Why Kindgi.