Foundations First

My son is wrapping up kindergarten. He has learned an extraordinary amount this year — letters, sounds, numbers, the discipline of sitting still long enough to listen to someone read him a story. None of it arrived in a single moment. There is no scene I can point to and say that was it. The foundations were built by stacking small things on top of small things, in a room where the adults around him had organised the conditions for that stacking to happen — a safe space, a teacher willing to slow down, a curriculum that knew what came next, the patience to step back when a kid needed to learn at a different pace.

That is not how the grassroots social purpose sector is being asked to learn AI. The sector is being handed the entire library on the first day and asked to read all of it by Tuesday. Of course it feels overwhelming. When the framing on offer is learn everything, we don't have time becomes the answer that comes back.

The reframe is small, and I think freeing. AI in your organisation is an organisational design problem. Every organisation I have worked with already knows how to do organisational design — that is most of what leadership is. The hard parts have been obfuscated — coding being the main one. All kinds of coding: database administration, getting disparate cloud applications to talk to each other, the endless spreadsheets and emails moving between teams. The technical skill required to do those things has been deliberately abstracted away by the tools themselves. What is left, once the obscuring is done, is something the sector is already fluent in: who reports to whom, what they're allowed to touch, what we expect them to deliver, and what happens when they get it wrong.

This post is the first in a longer series. The argument runs through it like a spine: start with the atomic unit, build the building block, stack the blocks. An atomic unit is the smallest piece of work that still does something useful on its own — a single Lego brick. A building block is a few atomic units stacked into something specific — a clear job description, a documented procedure, a tool with the right access. Stack the blocks and you have a working team. Nothing more complicated than that. The next post walks through what I have built in the last six months using exactly that logic — the visible proof. After that, two more — one on how leadership clears the runway for someone on staff to try, one on what day one looks like for the staff member who is doing the trying.

I have watched non-profits sit through the older version of this story a few times now. The IT consultant arrives. The CRM rollout begins. A year of change-management work compresses into three workshops. The staff resist. The board signs off the budget overrun. The system goes live missing the one function that mattered most. A few years pass. A different consultant arrives. The pattern repeats, slightly differently each time, but the shape is the same.

If you are positioning AI adoption as the next round of that, you are already in trouble. This is a different kind of project. The user interface is a conversation. The tool teaches itself in front of you. There is no week-long training course, because nobody knows the right shortcuts yet — including the people who built the thing. Tools that teach themselves require a different posture from the people deploying them, and the posture the sector has practiced for tech rollouts does not fit this work.

What is the right posture? It is the one the sector already uses for the rest of its work. AI participates in your organisation in the same way as a staff person does, or a volunteer, or a vendor. That is the role we are taking on, and the role the sector needs to learn to design for. The technical side becomes an afterthought once you accept that frame. Tools work by being asked questions. Their answers improve when the questions are precise. The judgement about what counts as a good answer stays with you. All of that is organisational design — the same kind your team already does, in the same vocabulary you already use, for the rest of your work.

A small translation table. Left column: what the AI world tends to call something. Middle column: what your organisation already calls the same thing. Right column: what the thing actually does in a working agent setup. You already speak most of the vocabulary.

It is the same exercise a kindergarten classroom runs on day one. Letters get sound names. Sound names get pictures. Pictures get words. The vocabulary translates because the underlying work is the same.

That table covers maybe 80 per cent of what a leadership team needs in order to read an agent setup intelligently. The remaining 20 per cent is jargon I would have to explain — and I would, in the same way I would explain deferred revenue to a new program manager: by walking through one example until it stops being mysterious. None of it is the hard part.

The harder skill, and the one the sector is already strong at, is organisational design. Limited span of control. Single owner. Transparent expectations. Segregation of duties — the principle that no one person should be able to both authorise and execute the same transaction. Documented procedures. Supervision arrangements that match the risk of the task. Those are the actual controls. Every one of them transfers, almost without modification, to how you set up a team of AI agents to do real work inside your organisation.

Those are the same controls a kindergarten classroom relies on — known adults, a daily routine, posted rules, regular check-ins, scissors that physically cannot cut skin.

You are already doing this with the humans on your team. The leap is small.

A small example, since this lands closest to home for many of the orgs I work with. A grassroots non-profit that has a bookkeeper but cannot afford an accountant has been carrying a quiet gap for years. The bookkeeper can keep the books accurate. The bookkeeper cannot, on their own, run the analysis a finance committee actually needs — trends, ratios, variance against sector benchmarks, gaps in the internal controls. With a junior assistant agent, properly scoped, that work becomes feasible. The agent reads the data, drafts the analysis, flags the gaps. A subject-matter expert validates anything material before it leaves the office. You cannot hold AI accountable, so the accountability stays with the bookkeeper and whoever signs off. But the bandwidth multiplies. The same pattern applies to administrators, fundraisers, project managers, HR staff — every core role with use cases currently managed on the side of someone's desk without the resources to deliver on the mandate.

The moving parts at the atomic-unit level. Each agent in my setup gets the same four files. Three are short, plain-language documents. The fourth is the same kind of optional reference document you might keep beside any role description — a tools-I-work-with cheat sheet.

Here is the directory for one of those agents — the chief agent I use to set up every new agent in my system. I will go into chief in much more detail in the next post; today I want you to see the shape of it.

Here is the opening of that agent's manifesto — the CLAUDE.md file at the top of the directory. Markdown (.md) is the file extension; for our purposes it is no different from a plain Word document. Use the Document toggle in the top right corner of the box to see it the way you would read a Word doc. It really is just a Word doc.

That is plain text. If your team can write a job description, your team can write this. The agent reads it at the start of every conversation. The behaviour shifts to match. The tool, which yesterday felt vague and slightly intimidating, is now a junior staff member with a clear remit and a list of things they are explicitly not allowed to touch.

I use Claude Code for most of this work, which is a coding tool that runs in the terminal — not the same product as the consumer chat at claude.ai. The file shape works across the major AI tools, though: Anthropic's Claude, OpenAI's Codex, Google's Antigravity (Gemini). The only meaningful difference: Claude reads CLAUDE.md, Gemini reads GEMINI.md, OpenAI reads AGENTS.md. The contents are the same plain text. Pick whichever tool your organisation already has a licence for.

In the next post I open chief up alongside three other agents — the social-media coordinator, the executive assistant, and the workflow-automation builder — and let you switch between them and read inside each one. You will see how four very different roles get the same four-file treatment and stay coherent across the whole team.

Whenever I explain this to a board or leadership team, five reasons come back for why now is not the right time. Four of them are reasonable. One of them is not. Open any card for the response.

You do not need to roll AI out across your organisation. You do not need to pilot it. You do not need a strategy document. You need to clear a small amount of capacity for one curious staff member to try something safely, inside a few clear guardrails, on work that is low-stakes if it goes wrong. Then you need to watch what they build.

That curious staff person — the bookkeeper running a research query, the fundraiser drafting outreach, the operations coordinator cleaning up a spreadsheet they have not had time to fix in two years — is the unit you are designing for. The agent is the junior assistant they did not know they were allowed to ask for. The building blocks are theirs to stack. Your job is to organise the conditions for the stacking to happen. Policy. Permissions. Time. Patience. A safe space to try, a curriculum the team agrees on, the discipline of routine — the same conditions, in a slightly different vocabulary, that the adults in my son's kindergarten organised for him this year.

If the sector decides this is organisational design and starts treating it that way, every grassroots organisation in Canada gets a junior research assistant, an administrator, and a first-rate junior programmer available to its staff at the cost of a discounted subscription. That is the window that has opened right now. I think it is freeing. I do not think it is overstated.

In the next post I walk through what I have built since I started taking this work seriously. The chief agent that builds the other agents. The social-media coordinator. The executive assistant. The workflow-automation builder. The mistakes — data structures dropped, social posts wiped, secrets exposed — and the controls that came out the other side. The five-layer safety pipeline that runs every time one of my agents generates an image. The actual sticker price: roughly a hundred dollars a month, plus a few cents per image generated. That is the proof that this is feasible. Today I wanted you to see the principle.