How I work with AI

A useful AI session should leave something I can use again: a document, a piece of working code, a decision recorded with its reasons. As I began using AI across more of my day, I wanted that work to carry over between sessions. An agent picking up a project should be able to find the context without asking me to explain everything from scratch.

I study at ESCP and work in account management at OVHcloud. Outside work, I build software and pursue independent research with AI assistance. My personal setup connects those longer projects with ordinary tasks such as managing appointments and keeping track of what I need to do.

How the system is organised

I give each part of the setup a defined role. Obsidian holds durable context. ChatGPT Work and Codex are where I ask an agent to investigate a question or produce something. Notion is the interface I have chosen for reading summaries, reviewing work and making decisions.

01 How I divide the work

Decisions Me

Set the task. Review the result. Approve the next action.

Reading & review Notion

The view I want to open: priorities, summaries and decisions.

↕ Brief & review

↕ Present & revise

Execution Work / Codex

Read context, use tools, produce an artifact and report what changed.

↕ Retrieve & update

↕ Read & act

Durable context Obsidian

Project notes, preferences, decisions and next steps in text files.

Source applications Mail, Drive, Calendar, GitHub

Documents and records, accessed through available connectors.

Everyday schedulingMe ↔ WhatsApp / Instinct ↔ appointments, calendar and to-do

The original records stay in their applications. Calendar events belong in the calendar, shared documents in Drive, and code in GitHub. Notes connect that material to the project: why it matters, what I have decided and what should happen next.

This separation makes the setup easier to maintain. I can change the model I use for a task while keeping the project’s files and history. It also gives me a way to judge a new tool: what part of the work will it make easier?

Context that survives the conversation

I use Obsidian as a text layer for agents. It holds project notes, preferences, decisions, sources and unfinished work. My intention is that agents read and maintain this material as they work, so I spend less time reconstructing context or managing a knowledge base myself.

Notion serves a different need. I want it to show the things that require my attention in a form I can read, edit and approve. This division between agent memory and my own interface is an explicit convention in my setup. Applying it consistently across workflows is still work in progress.

For research, a useful project note includes the question being investigated, the sources, the approaches already tried and the results that survived checking. It should make a failed attempt easy to find, so the next session does not repeat it. When a decision changes, I want the existing record updated; several competing summaries make it harder to know what is current.

I have also set up clipping into Obsidian to retain useful material. Another experiment was a morning brief combining mail, my personal calendar and at most three priorities. That scheduled brief is currently paused. Its format reflects a preference I apply more broadly: bring the important information together and make the next decision clear.

The tools I use

Different tasks benefit from different interfaces. I use a work session for something that needs investigation, files and several steps. For a quick appointment or reminder, I want a tool I can reach from my phone with very little effort.

ChatGPT Work / Codex

Research, code, experiments and document preparation, with access to the files and tools needed for the task.

Claude

An assistant in my work routine, including recurring tasks I have automated in that environment.

Instinct

An agent on WhatsApp that I use to book appointments, update my calendar and manage my to-do list.

NotebookLM

Reading and working through a defined set of source documents.

Google Drive

Shared documents and learning material that other people can read and contribute to.

Obsidian / Notion

Durable project context for agents, and a reading and review interface for me.

Connectors give an agent access to applications where the information already lives. MCP, the Model Context Protocol, is one way to make those tools available. With a working connection, an assistant can retrieve a document or inspect a repository directly, which saves me from copying the material into a chat.

I have also been testing LinkedIn and X integrations. Authentication and permissions have been practical obstacles, so I distinguish between having a connector installed and having a workflow I can rely on. Instinct is a recent addition, chosen because WhatsApp is already part of my day.

What I do with it

Handle everyday tasks

Instinct gives me a simple place to send scheduling requests and to-dos as they occur to me. The useful outcome is a calendar entry or task I can find again. For work that needs more context, I open Work or Codex and specify the result I need, the relevant files and the constraints.

I want an agent to finish with a clear account of what changed. If it prepared a document, there should be a document to open. If it changed code, it should explain what was checked and what remains unresolved. That makes it possible to review the work and decide what to do next.

Build shared learning material

I am helping two friends develop a working understanding of AI and the technology around it. We use shared Drive material to collect explanations as questions arise, covering topics such as agents, hooks, coding terminology and orchestration.

AI helps me develop the explanations, while I decide what the reader needs to understand. A useful explanation of a hook says when it runs, what it does and where it helps in a real workflow. Saving that explanation in a shared document makes it available to the group beyond the original conversation.

Carry out research I can check

The signed-difference-set project is my clearest public example. I used Codex for computational exploration, software engineering, organising artifacts and drafting. The project required explicit checks before a proposed result could become part of the evidence.

02 The research workflow

ExploreAI proposes a candidate

An idea, a program or a possible construction.

↓

CheckRun an explicit validator

Test the result against the stated mathematical conditions.

Fails or remains uncertainRevise the experiment

Keep the record; do not promote it to a result.

Passes the required checksPreserve the evidence

Save the artifact, code, inputs and verification record.

The repository contains constructions, independently implemented validators and exhaustive checks for the relevant finite cases. Failed searches and unchecked solver responses are excluded from the final evidence. The AI-use statement records what AI contributed and the limits of the verification.

This lets me delegate substantial work while keeping the standard of evidence explicit. An agent can propose an approach, implement it and run an experiment. I then need the artifact and the verification record to assess the claim. Someone else should be able to inspect the code and reproduce the relevant check.

Instructions I can reuse

I use skills and written project instructions to make useful procedures available across sessions. They describe where to find context, how to carry out the task, how to check it and where the result belongs. The rules I return to are straightforward:

• Read before continuing. Start with the existing project record, previous decisions and source files. Ask when a missing fact would change the action.
• Update the current record. Look for the document or task that already represents the work, and record useful changes there.
• Show what was completed. Point to the changed file, the check performed or the action taken. Keep unfinished steps visible.
• Respect the scope of permission. Outbound messages and invitations require my approval unless I have already authorised them.
• Keep claims tied to evidence. Distinguish an idea, an implementation and a result that has passed the required checks.
• Use my attention carefully. Consolidate updates, explain what matters and make the decision easy to understand.

These instructions also cover writing. I want explanations that connect one idea to the next, define unfamiliar terms and use examples where they help. A diagram should clarify a relationship that would otherwise take more effort to understand.

What I am improving next

I already use these tools for coding, research, documents and everyday administration. The next step is to make the connections between them more reliable. Obsidian and Notion have defined roles, but the diagram above is not a claim of continuous synchronisation. Some connectors are still being tested, and some automations have been paused.

I want to be able to see what an agent read, what it changed and what still needs me. For an appointment, that may be a correct calendar entry. For research, it may be a reproducibility package someone else can run. Those concrete outcomes are how I judge whether the system is improving the way I work.