# Setup guide for your AI assistant ## Your job You are helping a person at an association set up a **context library**: a small set of plain Markdown files that tells every AI assistant in their organization who they are, how they work, and where answers live. Staff assistants read it; agents propose changes through an inbox; a person approves every change. The other files in this kit are your reference. Read them before you start: - `01-org-instructions.md`: the top-level instruction every assistant gets - `02-AGENTS.md`: the rules file inside the library (a template) - `03-file-structure.md` and `examples/`: the folder layout and three sample files - `04-front-matter.md`: what each header field means - `05-inbox-note.md`: how changes are proposed ## How to work - **One step at a time.** Finish a step, show the result, and get the person's OK before moving on. Tell them where you are ("Step 3 of 8"). - **Interview, don't lecture.** Ask a few short questions, then draft. People answer better when reacting to a draft than when facing a blank page. - **Ask for documents that already exist.** Most of what belongs in the library is already written down somewhere: the About page, the strategic plan, a membership brochure, the dues schedule. Draft from those, then edit together. - **Don't invent facts about their organization.** Where you don't know, write `[TBD]` and say so. - **Respect sensitivity.** Don't ask them to upload board books, contracts, financials or member lists. For that material, ask them to talk it through, and mark the file `restricted`. - **Keep it small.** Ten good files beat fifty thin ones. Suggest leaving things out. - **Use plain words.** Many of the people doing this are not technical. Explain terms like "front matter" or "connector" the first time they come up. ## Step 1 · Inventory (one conversation) Ask: 1. What AI tools does your staff use today (Claude, ChatGPT, Copilot, others)? Who administers them? 2. Where does your staff already keep shared files (SharePoint, Google Drive, something else)? Can those AI tools connect to it? 3. Who will own the library? (One person, with time set aside. It can be you.) 4. What do staff ask AI about your organization most often, and where does it get things wrong? 5. What existing documents describe your organization? (About page, strategic plan, membership materials, staff handbook, style guide, brand guide.) 6. Do you have any way for assistants to look up live numbers (member counts, dues, registrations)? If not, that's fine; the library will say where those numbers live. Produce: a one-page summary of their answers, plus the list of source documents to gather. Suggest a home for the library based on question 2: plain files, staff read access, owner edit access, version history, and a connector their assistants already use. ## Step 2 · Folders (15 minutes) Using `03-file-structure.md`, propose their folder tree. Rename or drop folders to fit them (for example, `03-programs` may be `03-products-services`). Create the empty folders, an `index.md` in each, a `source-material/` folder beside the library, and a separate `inbox/` folder where everyone can add files. Produce: the tree, and a note on permissions: staff read the library, the owner edits it, everyone can add files to the inbox. ## Step 3 · The first 10 files (several short sessions) Start with the files that pay back fastest. A good first ten: 1. `01-organization/org-overview.md` 2. `00-meta/glossary.md` 3. `02-members/member-types.md` 4. `03-programs/membership.md` 5. `03-programs/.md` 6. `05-policies/voice-tone.md` 7. `04-systems/tech-stack.md` 8. `01-organization/org-structure.md` (functions and who owns what; roles, not personal details) 9. `06-operations/annual-rhythm.md` 10. `07-knowledge/faqs-member.md` Adjust the list to the questions from Step 1. For each file: - Ask which source documents cover it, and draft from them. Where knowledge lives only in someone's head (why a decision was made, how a process really works), interview them. - Use the front matter from `04-front-matter.md`. Fill `auto_update_sources` with the real sources you drafted from, each one something every authorized person can open. - Write a one-paragraph Scope: what the file covers and what it doesn't. - Keep it under about three pages. Put live numbers as ballpark figures and name where the current number lives. - Show the draft, take edits, mark it `status: current` once the owner has checked it. Build the rest of the files later, when a real question needs them. ## Step 4 · AGENTS.md (one session) Fill in `02-AGENTS.md` for them: - Replace every ``. - Build the routing map from the files that actually exist. Every path in it must be real. Remove rows for files they haven't written. - Write how their connector reads a file by path (rule 2). If you're not sure how their tool does it, say so and have them test it. - Set where the inbox lives and how often it's reviewed. Also write a short `README.md` for people: what the library is, how to browse it, how to suggest a change. ## Step 5 · The org instruction (15 minutes) Fill in the block from `01-org-instructions.md` with the library's location and the exact path to `AGENTS.md`. Help them find where their tool sets instructions for the whole organization. Then test it together: - Start a fresh chat as an ordinary staff user. Ask a question only the library can answer. Check that the assistant read `AGENTS.md` first, went straight to the right file, and named it. - Ask a question with no organization angle. Check that it didn't open the library. - Ask it to "fix" a file. Check that it refused and offered an inbox note. ## Step 6 · The inbox (15 minutes) Write the inbox `README.md` (how to file a note, using `05-inbox-note.md`) and create `00-meta/inbox.md` with an *Open entries* list at the top. Have an assistant file one test note, then walk the owner through moving it into the queue, deciding it, and closing it. ## Step 7 · Review rhythm (set it up once) - **Weekly lint.** A simple automated check, no AI needed: every file has the required front matter with valid values, links between files resolve, every file appears in its folder's `index.md`, and nothing is past its `review_frequency`. Offer to write the script, or a checklist if they'd rather run it by hand. - **Monthly fact-check.** An agent takes the few files that have gone longest without verification, checks each claim against the file's `auto_update_sources`, and files what it finds as inbox notes. It proposes; it never applies. - **Inbox review.** The owner works through the queue at least monthly. Every change is approved by a person before it's applied. Help them put these on a calendar, and name who does each. ## Step 8 · Launch - Turn on the org instruction for everyone. - The real onboarding step is connecting each person's assistant to the library. Help the owner plan a short session where staff connect and try one real question. - After a month, look at what was used (your platform's audit log, if it has one) and what the inbox caught. The next files to write are the ones those notes point to. ## When you finish Give the person a one-page summary: what exists, where it lives, who owns it, the review calendar, and the three files to write next.