Honest labels for AI-assisted correspondence

I get more messages than I can answer. Not more than I want to answer: more than one person physically can. Email, Slack, Google Chat, document comments, all of it arriving faster than any calendar allows for composing thoughtful replies. For years my choices were the standard two: answer a fraction well and let the rest go quiet, or answer everything briefly and badly. Silence costs relationships. Rushed replies cost them differently.

There has always been a third option, but until recently it was reserved for people who could afford staff.

How executives solved this before AI

Earlier in my career I worked with senior executives whose correspondence ran through a small personal staff, and I learned the system by watching it. It had levels, though nobody ever named them.

Some messages the executive wrote personally, start to finish. Some they dictated: a staff member typed, fixed the grammar, and sent, and the words were still entirely the executive's. Some got only a directive ("reply warmly, decline the date"), and the staff, who knew the executive's voice and preferences cold, composed the actual message. And some the executive never saw at all: staff read the incoming mail, handled it, and replied over the executive's name, trusted to know what the executive would have wanted.

None of this carried a disclaimer. In the offices I observed, it was treated as an ordinary delegation convention: the executive owned the relationships and set the direction, and the staff were a trained extension of their judgment, accountable for the message they sent.

AI agents now put that same system within reach of anyone. I've built mine on my own time, around two tools I run: Ragbot (ragbot.ai) and Ragenie (ragenie.ai). But the old system doesn't transfer unchanged, because two things are different.

What AI changes

First, some platforms tell on you. The Slack connector I use adds a visible "sent using" tag to agent-sent messages. My signature explains what the agent did.

Second, an AI agent today is not a trained chief of staff. A human staff member who drafted thousands of letters for one principal almost never invented a fact. An AI agent sometimes does. If a hallucinated detail goes out over my name, presented as my own writing, the damage doesn't stop at that message; it makes every real message of mine a little less trusted.

So unlike the executives I watched, I disclose. The design question is what the disclosure should say. My first attempt tracked process: which review workflow had run before the message went out. That collapsed under its own complexity, and it answered a question no recipient was asking. The signature needs to answer a simpler question: whose words are these?

Late-August 2026 update: The signature wording has changed; the examples below preserve the wording discussed here. The lanes still hold.

Three lanes

So the labels track authorship. Every message I send now travels in one of three lanes.

Manual: my words, my hands. I write the message or edit it until the words are mine, and I send it myself. AI may help me research, draft or polish. I read every word, decide what stays and take responsibility for it. I add no agent label.

Ragenie: my words, my agent's hands. I composed the message, dictated it, or edited a draft until the words were mine, and my agent refined and transmitted it. This is dictation, modernized. It carries one signature, always the same:

🧞‍♂️ I wrote this with my Ragenie

Ragbot: my direction, my agent's words. I told the agent what to do (a specific instruction, or standing rules for a class of messages like scheduling and acknowledgments), and it wrote the words. This is the directive level of the old staff system. The signature says so plainly:

🤖 my Ragbot handled this one for me — I read every reply

The loop always closes with me.

After two or three messages, the legend teaches itself. No marker: words I own and send myself. Genie: my words, delivered by my agent. Robot: my agent, working under my direction. A recipient never has to parse fine print to know what they're holding.

The rule that keeps it honest

One rule holds the system together: a lane can only claim what actually happened, and when in doubt, claim less.

The genie signature makes a strong claim, I wrote this, so it is only available when that claim is true. There is no "mostly wrote it" genie variant, and no version of it for messages I haven't read. I know because I tried to write one, and every attempt came out self-contradictory. That failure was the design insight: you cannot phrase your way into claiming authorship of words you never saw. If the words aren't mine, the message belongs to the robot lane, whose claim (my direction, not my authorship) is honest at any level of my involvement.

The labels are backed by a workflow with three requirements: read the entire thread history before composing a reply, review the draft for voice and AI slop, and recheck facts at send time because approval of text is not approval of a changed world. A fail-closed hook requires a fresh record of those checks for the exact message and applies its language checks before allowing a send. The machinery serves the labels, not the other way around.

I've open-sourced the general system (the lanes, the persona registry that lets you brand your own equivalents of Ragbot and Ragenie under your own names, the gates) as synthesis-agent-correspondence in my synthesis-skills repository. The specific brands and signature wording stay in a small private config, which is the part that should be yours anyway.

The executives I watched had something most of us never did: enough trusted hands that no message worth answering had to go unanswered. I want an honest way to be present in more of my correspondence than my hands alone allow, with labels that tell every recipient how present I was.

Send permission is separate from labelling. My configuration requires me to see the exact text of every agent-lane message before it goes out. The published synthesis-agent-correspondence skill carries the lane model; what any one person permits lives in their own private configuration, so set that part yourself, and expect to tighten it after the first message you wish you had read first.