Recently, when I talk to dots, it has become possible to request work from both Codex and Claude Code.
I have made ChatGPT's dots the command center for my work.
I speak about what I want to do.
dots issues instructions to Codex or Claude Code according to the role division I provided in advance.
Then, it returns the results to me in an easy-to-understand format.
The role divisions for each, the content of the work, how to proceed, and the necessary data are all stored in Obsidian.
I use several AIs, but the only one I need to talk to is dots.
This is the biggest change in my current usage method.
Instead of just asking each AI to do tasks, AI now intervenes even in the process of deciding who to ask and how.
I Just Need to Talk to dots
Until recently, whenever I used AI, I had to decide myself which one to ask every time.
"Should this task go to Codex?"
"Should this one go to Claude Code?"
"I want to involve another AI midway, but how should I communicate that?"
Separate from the actual work content, logistical steps for requesting the AI would arise.
In my current environment, I leave this sorting to dots.
At the point of speaking, I don't need to create separate instructions for Codex and Claude Code myself. I tell dots what I want to do, and it acts according to the pre-assigned roles.
Between me and dots, we simply consult normally.
Beyond that, dots issues instructions to each respective AI.
I use Codex and Claude Code for their specific roles.
I decided their responsibilities in advance and consolidated the consultation window into dots.
If my logistical burden increased every time I added a new AI, it wouldn't be easier; I'd just get busier again...
Rather than how many AIs you use, it's about how much logistics you carry yourself.
I realized that this aspect can also be changed.
Just Assigning Roles Doesn't Convey Work Content
To have dots sort out work, information needed by the assigned recipients is also required.
Who handles what is different from how the work is actually done.
What I store in Obsidian isn't just the role division between Codex and Claude Code.
The content of each task.
The procedure for that task.
The data required to proceed.
All these things are also in the same Obsidian vault.
Even if the person in charge is decided, if they aren't given "what to look at and how to proceed," they will have to ask me.
For example, if the task is writing text, you need the target audience and writing rules.
If the task involves numbers, you need the source data.
These are examples of dividing work types, but neither can proceed based solely on "who to ask."
In my environment, I also store publishing rules and operational procedures in Obsidian.
This ensures the AI responsible for that task can proceed using the necessary information.
On the surface, the consultation partner is dots.
Behind the scenes, the information supporting the work is in Obsidian.
Even if I change where I consult, I haven't structured it so that I have to recreate the work explanation from scratch every time.
I am using the information I've accumulated in Obsidian for my current request method as well.
dots' Job Includes Making Sure I Understand the Response
Another important aspect of this usage method is the response.
dots doesn't just issue instructions to each AI; it also returns results to me in a way that is "easy to understand."
There is a time cost for the requester to read responses.
Even if work progresses, if the burden of deciphering explanations increases there, the receiver becomes busy again.
Even if details are written, if I don't know what to do next after reading them, I end up having to ask more questions.
For example, in a report, it's easier to receive if you can see what is finished, what remains, and what decisions I need to make.
Depending on whether you want to know the detailed background or just the key points to move forward, the desired format changes.
Therefore, I want to consider not just how to request work, but also how to receive responses.
Going out: dots takes my input and converts it into instructions for each AI.
Returning: dots takes results from each AI and returns them to me in an understandable format.
dots handles this round trip.
Being able to request complex tasks and receiving responses I can understand—both are necessary to proceed to the next consultation or task.
━━━━━━━━━━━━
For those who use AI but find publishing difficult.
Stop starting from zero every time by placing your materials in one location.
I started "Minami's Obsidian x AI Agent Implementation Log."
You can read it for free ↓
━━━━━━━━━━━━
Start with Work Content, Not AI Names
If you want to try a similar setup, instead of connecting multiple AIs from the start, pick one task you usually request to organize your thoughts.
For example, if the task is creating X drafts, just writing "create text" leaves the information passed to the assignee unclear.
What kind of publishing are you doing?
What materials are being used?
How far do you want it created before returning?
When these are written down, you can see what needs to be passed when separating the consultation window from the working assignee.
To organize work descriptions, you can use a framework like this:
1Task to request:2AI responsible:3Work content and procedure:4Location of notes/data to use:5Things to clarify in the response to me:
Store this description in Obsidian.
It feels like writing down not just who to ask, but what the assignee uses, how they proceed, and in what format they return results.
This is a memo to organize the explanation when passing work to the command center.
If the work content is visible beforehand, "what I want to entrust to the command center" also becomes concrete.
Instead of starting by lining up AI names, think from the work you are constantly connecting.
I believe this makes it easier to create a form that fits your own publishing.
Conclusion
If you are using AI but tired from deciding who to ask, how to communicate, and how to read responses, check if you are carrying everything yourself.
You don't need to always be the liaison to master AI.
I now talk to dots, leaving the instructions to Codex/Claude Code and the responses back to me to dots.
If you try today, pick one task you usually request from AI and write it out in Obsidian using the framework above.
Consultation partner: dots. Work content/procedure: Obsidian.
I am proceeding with multiple AIs using this combination.
Thank you for reading until the end!
Announcement 📢
I run an Open Chat!
Under the plan "Explode Obsidian Users!", I share real examples and failure stories of Obsidian x AI Agents that aren't posted on X.
I'm also distributing gifts, so please take a peek if interested ↓
