What is NicAI

NicAI is my personal AI layer: a set of agents, skills, tools and data, built around my own knowledge and my own way of working. It researches, drafts, builds and maintains things for me, at work and at home. I review what it produces, fix it, and feed the fix back so the next run is better.

It isn't one app or one model. It's a system I assemble and change all the time, with Claude Code at the centre today, and plain files underneath so any model can use them.

This section of the notebook documents the parts of that system, one note per part.

The goal

I described it in July 2025, in my first public post about NicAI:

• A portfolio of digital workers that use my knowledge and my style, from how I write to how I do things, and that can take on a wide range of tasks, eventually orchestrating themselves.
• A flywheel that learns every day. Every correction becomes a rule, a skill or a better data source.
• Privacy first. My data stays in formats and places I control.

The headline was "I'm planning to replace myself with AI". The point was the opposite: replace as many tasks as possible, so I can do more of the work only I can do.

How it started

The 2022 note was mostly a list of repositories to explore. The first working version, in early 2024, ran on OpenAI APIs, my own vector database and a lot of Python scripts. It went through many iterations after that. What changed most is the approach, not the models: from one assistant on top of a vector database to many small, specific skills on top of plain files.

How it works today

NicAI has 6 layers. Each one gets its own note in this section over time.

A few design choices hold it together:

• Plain files over platforms. Knowledge lives in Markdown, JSON and SQLite. Swapping the model is a line of code, and nothing is locked into a vendor.
• One skill, one job. Small skills are easier to test, fix and combine than one big prompt.
• Review loops for high stakes. The Council runs an artifact past a panel of expert agents that critique and score it until they converge.
• Scheduled upkeep. 4 nightly jobs refresh the wiki, the image index and the internal documentation, so the agents start each day with current data.
• Always a first draft. 80/20 quality on volume tasks, full rigour on deliverables that matter, and my eyes on everything before it goes out.

Where it stands

23 Sep 2026

What works:

• Research and first drafts at volume: meeting briefs, outreach briefs, one-pagers, decks, emails and notes.
• Upkeep: knowledge bases, databases, this site's publishing pipeline, and my own IT and home setup.
• Parallel work: several agents on separate tasks at once, with me reviewing and steering.
• This notebook: 103 of its 1,288 notes were researched and written by NicAI, each marked with a signature at the end.

In July 2025 I wrote that hours invested exceeded hours saved. By May 2026 a Saturday with 6 agents delivered what would have taken me weeks before. The investment is paying back now, but it still takes real time to build and maintain.

What isn't there yet:

• My voice. The drafts follow my rules, but they don't yet sound like me without editing. The next step is a small local model fine-tuned on my own writing (see Fine-tuning my own model (LLM): NicAI ). So far it's a plan, not a model.
• Self-orchestration. I still decide which agent does what, and I review every result. The agents don't yet plan and hand over work between themselves.
• Local-first. The privacy goal points to local models, but the heavy work still runs on frontier models in the cloud. Local models run on a Mac Studio (M1 Ultra, 128 GB) through Ollama, for experiments so far.
• Some things stay manual by choice. Outreach to people goes out by hand. NicAI prepares it, I send it.

Principles

• Document your own knowledge. Most of the effort goes into capturing what I know, do and decide in structured files. Models are a commodity; that data is the edge.
• Every correction feeds back. An edit I make once should become a rule, so I never make it twice.
• Model-agnostic. Each task gets the model that suits it.
• Human in the loop. NicAI drafts and builds. I decide, edit and send.
• Be open about it. Anything NicAI wrote in public says so.

What's next

• Fine-tune and blind-test the voice model against the current Claude plus persona-spec setup, and keep it only if it wins.
• Let agents hand work to each other on routine chains, with me approving at the end instead of at every step.
• Move more of the private work to local models as they get good enough.
• Document each layer in this section, one note at a time.

In this section

How it's built:

• The folders and the task queue: The NicAI workspace
• The instruction files, hooks, memory and permissions: My Claude Code setup
• Every skill, 1 line each: The NicAI skills catalogue
• The Python wrappers behind them: NicAI shared tools
• The databases and the nightly jobs: The NicAI data layer
• How NicAI writes like me: The NicAI voice stack

What it does:

• A day with 6 agents: Working with agents
• The review loop: The Council
• The nightly wiki: LLM Wiki
• How this site gets published: The notes pipeline
• The conversational avatar of me: My video twin
• The share space for what NicAI builds: nicailab.com

What I learnt and what's next:

• Time, money and what failed: NicAI: costs and lessons
• The voice model plan: Fine-tuning my own model (LLM): NicAI
• The writing tells NicAI avoids: The New Em-Dash: 9 Tells of AI Writing
• Common mistakes when building an AI OS: Setting Up an AI OS: 6 Mistakes and Their Fixes
• How this site is built: Building this site: notes.nicolasdeville.com