My AI Team Debated My Ideas at 5am While I Slept
A tour of the personal intelligence system I built to capture, research, and evaluate ideas overnight
Most of the conversation about AI right now is about products. What you can build, what markets you can reach, what revenue you can generate. That’s fine. But I don’t hear nearly enough about something I find more immediately useful: what you can build for yourself.
I’ve spent the last few months building a personalized intelligence system that runs quietly in the background of my life. It captures my ideas the moment they surface. It researches them overnight. By 5am, a team of agents has deliberated on each one and rendered a verdict while I was asleep. None of it is for sale. All of it makes me sharper.
Here’s how I built it, and why the architecture matters more than any individual tool.
---
## The foundation: ~/.dna/
Everything lives in a single directory on my machine: `~/.dna/`. I call it “dna” because this is where my personal tools save their output. All of it, in one place.
The pattern is simple. Every tool saves to `~/.dna/<tool-name>/`. Every tool has a `README.md` that acts as an index, a table of contents for everything that tool has ever produced. The directory structure is always the same:
```
~/.dna/<tool-name>/<session-name>/<output-file>.md
```
That consistency is what makes everything else possible.
---
## The idea tool
Earlier this year I built a simple Claude Code command called `/idea`. Any time during a working session, when a thought surfaces I don’t want to lose, I type `/idea my amazing idea is...` and the system generates a new file, writes the raw idea, and asks me one question: quick capture or brainstorm?
Quick capture means done. Back to what I was doing.
Brainstorm means it pulls the idea out of me one question at a time, writing as we go. The output lands in `~/.dna/ideas/<idea-name>/idea.md`. The README updates automatically.
I love this command. No idea gets lost to the friction of switching context.
Each idea has a status: `raw`, `exploring`, or `completed`. Completed means I’ve fully fleshed it out. That status unlocks the pipeline.
---
## The market research pipeline
Every night around 11pm, a market research agent runs against any idea marked `completed`. It spins up parallel agents to research thirteen areas at the same time:
Buyers Journey, Community Mapping, Competitive Analysis, Demand Signals, Investment Landscape, Market Research, Market Sizing, Regulatory Scan, Review Mining, Search Demand, ICP Construction, Mom Test, Trends Research
Each agent writes its own report. A synthesizing agent reads all thirteen and writes a master report. Everything saves alongside the idea.
The research isn’t meant to replace the work you’d do if you were serious about an idea. It’s meant to give you enough signal to decide whether to be serious. Is there a market here? Is it saturated? Are real people actually feeling this pain?
---
## The council
This is the part that surprised me most.
Every morning at 5am, a council of three agents sits down with any idea that has completed market research. They each read the idea, the market research, and an `identity.md` file. That file is a detailed profile of who I am, what I’m capable of, what I’m not, and what kinds of businesses actually fit me.
Then they deliberate. Without me. I’m still asleep.
**The Skeptic** looks for failure modes, untested assumptions, and conditions under which the idea fails.
**The Advocate** builds the strongest case for why this idea works, what market force makes it viable, and what specific capability match makes this the right bet.
**The Pragmatist** evaluates what actually ships, whether the build estimate holds, and where the plan breaks between paper and production.
Four rounds. Then a Summary Agent reads everything, including the private thoughts each agent generated during deliberation, and produces a full `deliberation.md` and a clean `council-decision.md`.
The council gives me each agent’s vote, a personal fit score, the risks, and action items. Some of the feedback has been sharp: “Anthropic could release this, which means you potentially have no moat.” Some has been practical: “Can you sell this without building anything first to test whether customers will actually pay?”
The personal fit score is what I find most valuable. The market research tells me if there’s demand. The council tells me if this idea is right for me, given who I am, my actual skills, my risk tolerance. That’s a harder question. The council asks it every time.
---
## Memory: short-term and long-term
One pattern runs through all these tools: markdown files are short-term memory, SQLite databases are long-term memory.
Markdown is fast, readable, and easy for agents to work with. SQLite holds the conversation logs, deliberation transcripts, the data that needs to persist and be queryable over time. One database per tool, when I need it.
A Memory Keeper agent runs nightly. Its job is to look at each tool’s files and decide what should move to long-term storage.
---
## The brain
Once all these tools were saving consistently to `~/.dna/`, a natural question came up: how do I query across all of them?
The answer is a brain command built on a single file: `~/.dna/map.md`. The map is simple. A description of each tool, where its index lives, and what questions it can answer. Agents read the map at startup and fetch what they need. The brain doesn’t copy everything into one place. It knows where things live.
Think of it like DNS. The map is the root server. Each tool’s README is its own nameserver. You don’t store every address in one file, you know who to ask.
This means I can ask the brain whether an idea lines up with my current goals. I can pull patterns across my coaching sessions and my idea history. The brain can talk to any of the specialized agents across my tools and route to whoever can answer the question. One of my favorite things is pulling information across tools and having a team of psychologist agents analyze my thinking. It’s genuinely fun.
---
## Where to start
If this inspires you to build something similar, start with `identity.md`.
Ask an LLM to interview you. Not “what are your goals,” a real interview. It needs to understand who you actually are: what you’re good at, where you need work, what kind of work pulls you in and what drains you. Save it to `~/.dna/identity.md`.
Then build an idea command. Capture everything. Keep the structure simple and consistent. Add tools when the need is obvious.
The directory structure and the index files are the architecture. Everything else builds on top. Once you have that foundation, you can string tools together and start getting insights none of them could give you alone.