TL;DR: AmpereBrain builds an AI sidecar for KiCad — software that does the grunt work of hardware design so engineers can focus on decisions. We chose to run the entire company the same way: one Git-backed Markdown vault, strategy as plain text, AI agents as the org chart. Not because it’s fashionable. Because a company that sells “AI does the work” can’t run its own shop on data AI can’t read.
Don’t use what you’re used to
New companies get a gift existing companies don’t: a clean start. No Google Drive. No final_v7_REAL.docx. No folders named after the intern who left.
The instinct is to copy the stack you used at your last job — Notion for docs, Slack for chat, Google Drive for files, Jira for tasks, a CRM you’ll resent by month three. Day one, before you’ve shipped anything, you’re already renting your own data back from five vendors. Each new hire costs a seat. Each export is a support ticket. Each tool is a silo AI can’t reach without an API key and a prayer.
The real cost isn’t the subscriptions. It’s that on day one, no machine can reason about your company. Your strategy lives in a Notion page AI reads as a blur of blocks. Your roadmap lives in Jira behind a login. Your customer notes live in a CRM that charges you to leave.
If you’re competing against AI-native companies, you won’t follow the cadence.
The philosophy: the company is the product
AmpereBrain’s product philosophy is four words: intent first, checkpoints always.
The engineer defines intent — the specs, the noise floor, the power budget. The AI executes the grunt work — datasheet hunting, library linking, simulation setup, version archiving. The engineer reviews at checkpoints: this component doesn’t meet your spec, approve or redirect. No black box. No blind trust. The human owns every decision; the AI owns every step between decisions.
We run the company on the exact same loop.
- Intent lives in
Project_Plan.md— the OKRs, the phases, what matters this quarter. - Execution is delegated to AI agents — the advisor, the product strategist, the community scout, the content marketer. They draft, research, and do outreach.
- Checkpoints are where I review: the strategy call, the message edit, the pricing decision.
Dogfooding gets called a marketing line. It isn’t. It’s the only way the team stays honest. If we sell “AI does the grunt work,” we can’t run the company on grunt work. If we sell local-first and data sovereignty, we can’t run the company on SaaS hostages. The company has to be the proof.
Why it has to be plain text
An AI agent pointed at a Notion workspace sees a rendering layer, not knowledge. Pointed at a Markdown vault, it sees the knowledge itself.
That’s the whole argument. In software engineering there’s a principle — KISS, Keep It Simple, Stupid. You want your data understandable, not “smart.” The moment you bolt on an external indexing service, a vector database, a “knowledge layer,” you’ve created two sources of truth: the one you work in and the one the AI works over. They drift. They disagree. That’s not a feature; that’s a bug factory.
So: simple text files. Small scripts when you need to make them look good — a Marp deck for the KiCon talk, a Pandoc PDF for the investor note. Nothing else. The vault is the knowledge. The AI reads the same files you do.
The structure: one source of truth, agents as the org chart
The vault has a few non-negotiables.
One master document. Project_Plan.md is the source of truth. Every task maps to a Phase (0–4) and a Key Result. If a piece of work isn’t in the plan, it doesn’t exist. There is no second Project_Plan_FINAL.md. There is no “the real roadmap is in my head.” One file. One truth.
Agents are the team before there’s a team. AGENTS.md defines a roster of specialized agents — startup-advisor, product-strategist, community-scout, technical-educator, growth-catalyst, and a dozen more. Each has a role, an archetype, and a remit. That is the org chart. When there’s no head of marketing, the content-marketer agent holds the seat. When there’s no CFO, the financial-steward agent does. The company has a full executive team’s worth of specialist work available the day the repo is initialized — staffed by models, governed by rules I wrote.
Numbered, self-describing folders. No Misc/. No Stuff/. Every path tells the AI what’s inside before it opens a file. Grep respects folder boundaries; so does access control.
Everything is Git-backed. Full history. Offline-first. No vendor can revoke my access to my own strategy. (If Git is new to you: it’s a versioning system that keeps the complete history of your text files and syncs them through a host like GitHub or Forgejo, where you decide who can reach what.)
The payoff: what day one looks like
This is where it stops being philosophy and starts being unfair advantage.
One founder, plus the agent roster, produces the work of a much larger team. A strategy review is “point the advisor at Project_Plan.md.” Competitive analysis is “point the community-scout at the ICP and the market.” Launch copy is “point the growth-catalyst at the value proposition and the launch plan.” Every output is grounded in the actual company — because the actual company is the only thing the AI can read.
No context is lost between tasks, because the context never lived in a chat. It lived in the vault. The advisor agent and the product-strategist agent read the same Project_Plan.md. They disagree out loud, on the page, where I can see it.
Try that with five SaaS tabs open. You can’t. The knowledge is scattered; the AI is blind; and you’re paying $200 a month for the privilege.
How long it took
People hear “company operating system” and picture a months-long IT project. It isn’t.
- The scaffold: an afternoon. One master repo, the numbered folders,
AGENTS.md,Project_Plan.md,.gitignore. The structure is a template, not a research project. - The hard part: the discipline. Keeping strategy in one file. Writing the OKRs before chasing tasks. Refusing the impulse to open a Notion page “just for notes.” That is the real investment — not the setup, the habit of using your agents first.
A fresh start has no 5,548-file vault to rescue (see the StepUp Horse migration case study for what that costs). We never created the debt. That’s the entire point of starting AI-native: the cheapest migration is the one you never do.
The difference from a migration
This is worth stating plainly, because the two paths get conflated.
| Migration (existing company) | Fresh start (new company) | |
|---|---|---|
| Starting state | 5,000 files in folders named after people | Nothing |
| Main risk | Losing files, breaking access | None — there’s nothing to break |
| Main cost | Conversion, validation, re-education | Discipline and habit |
| Time | Days of work, weeks of habit change | An afternoon of setup, ongoing habit |
If you already have a company, migrate. If you’re starting one, don’t create the mess you’ll later have to migrate. Start clean. The cost of doing it right on day one is a rounding error next to the cost of fixing it on day 500.
What’s next
Next up: Start an AI-Native Company — set up your folder structure, file organization, and team access. The how-to this case study argues for.
The full path after that:
- General Setup — Git vs Google Drive, Run your AI Agents, AGENTS.md Template
- Skills — Markdown Essentials, Git Essentials, Train Your Team
- Tooling — CRM, PDFs, Decks, Social
One vault. One source of truth. One company the AI can operate from the day it exists.
Your data. Your rules. Let’s write it that way. By Charles Henri Gayot.