I just redesigned this website, which is a sentence that sounds like I spent a month lovingly handcrafting components in a code editor. That is not what happened.
I wrote no code. I architected the thing, governed the work, argued with a coding agent, reviewed what it produced, and kept pushing it toward the system I wanted. The first version was almost one-shot designed, including the seed of the Direct Message Reference idea. Then I became more interested in the governance layer than the site itself, which is exactly the sort of detour that explains both my personality and why I have to write my ideas out first.
The agent kept losing context because English leaves too many valid paths open. If I don’t constrain it, it will explore, blend approaches, and drift.
I start in English to define the problem, but that’s just the entry point. From there, I collapse it into a set of explicit decisions, high-confidence scoring, and ultimately hard pass/fail checks. That becomes the standard.
The problem defines the solution plane. The standards confine the context. The deterministic guardrails handle pass/fail. Together, that creates the loop. The agent produces a solution, it gets tested against the guardrails, and if it fails, it refactors back inside the allowed space. That loop keeps it moving toward the goal without inventing a new one.
It was fun to work through this loop. It was also slower than the fantasy version where I say “make website” and then go outside like a well-adjusted person.
The DMR feature came out of that process. I removed the normal comment box because I am not convinced the normal comment box is optimized for anything useful.
That sounds dramatic. It is not. It is a small experiment on a personal website with a tiny audience, which is exactly the right place to test something weird without pretending it is a movement.
The current design is simple: if you want to respond to something I wrote, you can highlight the passage or use the floating + button. The site creates a Direct Message Reference (DMR): article, passage, selection if there is one, and enough context that the message does not arrive as a vague “about your post” drive-by. Then it opens a DMR path through email, WhatsApp, or Telegram.
No public thread. No like button. No “share this to social media and please feed the machine” button. If the idea sparked something real, talk to me directly.
The Theory
I write to think.
That is not a brand statement. It is the actual mechanism. Most of solving a problem is forcing the problem into words. Once the problem is written down, the weak parts become easier to see, and the quantitative parts reveal themselves as guardrails. The gaps stop hiding behind vibes.
I write for myself. But if the writing sparks a real response from someone else, I am interested in that. Especially if the response says the idea is wrong.
That is the useful part.
Disagreement can make a thesis more rigorous. A good objection is a gift. A public dunk is usually just someone doing free labour for an algorithm.
I am not interested in rage posting. I am not interested in emotional trap algorithms. I am not interested in compressing a hard point into a tiny social post and then acting surprised when the result is mostly heat.
So the theory is:
If the response path is direct, contextual, and slightly higher friction than a public reaction box, the quality of the response may improve.
Maybe that is wrong. That is why this is an experiment.
The Test
The test is whether Direct Message References create better conversations than a public comment board or social sharing.
The site does a few things:
- It lets you respond to a specific passage, not just the whole article.
- It attaches enough context that I know what you are responding to.
- It routes the response through a private channel.
- It allows people, agents, and mixed workflows to respond.
- It avoids creating a public performance surface.
I considered a Discord server. I still might do that one day. Right now, I do not think there is enough signal to justify administering a room that mostly exists so three people and a notification bot can stare at each other.
The smaller test is better: DMRs first, community infrastructure later if the conversations earn it.
The Agentic Layer
There is some agentic code around the DMR flow.
At a high level, the site creates a DMR intent when you choose a channel. That intent stores the article, the passage, the selected text if there is any, and some lightweight metadata about how the interaction happened.
The point is not surveillance. The point is context.
“I disagree with this” is not very useful if I have to guess what “this” means. “I disagree with this paragraph in this article, and here is why” is much more useful.
The DMR path is also built to support agent participation. If an LLM or agent wants to respond to a piece, I want that to be allowed. I am curious if the agent will disclose itself. The checkbox exists for that reason, otherwise, it’s of no real consequence.
The Confidence Algorithm
The site has a lightweight confidence layer that tries to infer whether an interaction looks more like a person, automation, an agent, or unknown.
It is not a truth machine. It is a rough signal. It can be wrong. It will be wrong.
Sometimes a human looks automated. Sometimes an agent looks thoughtful. Sometimes both are true because a person is using an agent well.
That matters because I do not think the interesting future is “humans over here, agents over there.” The interesting future is mixed conversation: people thinking with infinite intelligence, agents extending people, and systems that need to be honest enough about their role that the conversation does not become fake by default.
The confidence layer is there to help me understand the shape of the interaction, not to decide who is allowed to talk.
How I Respond
If you send a message, I may respond directly.
Sometimes an agent may help first. The agent is intentionally small and narrow. It is mostly trained or grounded around my writing, including material that is public and material that is not. It is not some grand autonomous personality. It is more like a mildly useful intern who has read a lot of my notes and still absolutely needs supervision.
Neither of us is always online.
The model behind the agent and the agent itself run on some on-premise infrastructure, which is the dignified way of saying a small machine I use for experiments, proof-of-concepts, and whatever else I am currently breaking. I may reboot it, repurpose it, break the agent, forget to restart the agent, or decide the machine has a more important job that day. There are no uptime guarantees here. This is not a production system. It is not a five-nines service. It is a lab feature on a personal site, so if the agent is offline, slow, confused, or missing entirely, congratulations, you have discovered the lab.
You can try to trick it. I do not mind. If it responds badly, that is useful too.
I watch the messages. The purpose of the system is not to avoid people. It is to create a better first exchange so the real conversation has more context and less performative nonsense.
The Channels
Email and WhatsApp are the cleanest paths.
Telegram is a little more annoying because the share flow cannot reliably force the recipient. To make Telegram work as a DMR path, it needs to go through a bot so the passage context can be attached and the message can be routed correctly.
That is not because Telegram is bad. Telegram is great. It is just not built around this exact “create a Direct Message Reference from a passage on a static website and route it privately” use case.
What Would Make This Work
This experiment works if it produces fewer responses but better ones.
I do not need a public count. I do not need engagement theatre. I want the thing that public comment systems often pretend to want: thoughtful disagreement, useful corrections, surprising extensions, and the occasional “you are overthinking this” from someone who has earned the right to say it.
The DMR success criteria are simple:
- Did the response teach me something?
- Did it make the original idea more rigorous?
- Did it create a conversation worth continuing?
- Did the DMR path reduce low-quality performance behavior?
- Did agents contribute signal without pretending to be people?
If yes, the experiment is worth keeping.
If no, I will change it.
The Larger Question
The larger question is whether personal publishing should keep copying the social web.
The social web is optimized for distribution. That is useful, but it comes with a tax: outrage, compression, incentives that reward certainty over thought, and the constant pressure to turn every idea into a little engagement pellet.
I do not want this site to be that.
This site is a notebook, a lab, and a place to put ideas where I can find them again. The Direct Message Reference experiment is one more part of that structure.
If people are interested, I may write a deeper technical post on the agent itself: model choice, prompts, retrieval, confidence scoring, handoff behavior, and where the whole thing is embarrassingly dumb.
For now, the important part is simpler.
If something here makes you think, message me.
If you think I am wrong, even better.