--- title: Building context description: Structure and upload the context you already have, so every agent on the team can read it. --- Every team already has most of the material for its context layer. It is just scattered: in a website's wording, in a handbook nobody reads twice, in the instruction files you keep for your own AI tool, in a spreadsheet someone updates on Fridays. The five prompts on this page each pick up one of those sources and turn it into something every agent on the team can read. They are also five answers to the question "what would I even put in here?" Read the short introductions and start with the one that sounds most like your situation. Each stands on its own, and the order below runs from the least you need ready to the most. Each prompt runs a short conversation. The agent looks at what you already have, asks you a question or two, shows you what it intends to save, and saves it once you agree. You stay in control of every word that enters the workspace. ## Define your company language Your team's own words, defined once. When a skill or a document uses one of them, every agent reads the same meaning, because the definition travels with the word instead of being guessed afresh by each tool. The prompt works from what the agent remembers about your company and what it can read of your website. ```text Foundation setup: Define your company language Help me put my company's own words into Foundation, our shared context layer, so every AI tool on the team means the same thing by them. A term is a word or phrase we use in our own way: a product name, a customer stage, internal shorthand, anything a new starter has to ask about. First, check you can write to Foundation. If you have no create_entity tool, say so in one sentence and stop: no plan, no questions, no request for material. I'll paste this where Foundation is connected. Before proposing anything, look at what Foundation already holds with list_entities, so you never create a near-duplicate and can reuse its tags. Look before you ask: what you remember about my company, our website if you can reach it, anything I paste. Collect the words we use in our own sense, with the definition where the material gives one. Use only what you read or what I tell you. If you have a word but not what we mean by it, ask me; never guess a definition, and never fill the list with words other companies use. Then ask me one question: which words does a new starter have to ask about, and did I miss any? If I say I'm not sure, show me what you found and ask which of it counts. If you found nothing and I give you nothing, say so and stop. Propose up to eight as a table, exactly as they will be saved: the word, its one-sentence definition, an example of it in use, where it came from. Then ask "Shall I add these, or change anything first?" and wait. On my yes, add each one to Foundation as a term: create_entity for a text marked as a term, with the one-sentence definition as its description, then write_text with the definition and the example from the table, nothing more. Reuse tags Foundation already has. When all are in, tell me what was added in one line. Never define a word I did not confirm. ``` ## Share your AI's memory with the team What your own AI tool already knows about the company becomes the team's. The prompt sorts it into ways of doing things, words and facts, and leaves out anything personal. ```text Foundation setup: Share your AI's memory with the team Help me move what you already know about my company into Foundation, our shared context layer, so the whole team's AI tools know it too, not just mine. First, check you can write to Foundation. If you have no create_entity tool, say so in one sentence and stop: no plan, no questions, no request for material. I'll paste this where Foundation is connected. Before proposing anything, look at what Foundation already holds with list_entities, so you never create a near-duplicate and can reuse its tags. Look before you ask: what you remember about me and my company, and any instruction files you can read. If you cannot see any memory from here, say so and ask me to paste what my AI tool has saved about the company. Use only what you find or what I paste; never make something up to have something to propose. If there is nothing, say so and stop. Sort what you find into three piles, in plain words: ways of doing things the team should follow, words we use in our own sense, and facts about the company. Leave out anything personal, and anything that is my own setup rather than the team's; say what you left out. Then ask me one question: which of these should a new coworker's AI already know on their first day? Propose up to five as a table, exactly as they will be saved: name, which pile it is in, the content in full, where you found it. Then ask "Shall I add these, or change anything first?" and wait. On my yes, add each one to Foundation: create_entity for a text (marked as a skill for a way of doing things, as a term for a word, plain for a fact), with a one-line description written for an AI saying when to reach for it, then write_text with the content from the table, nothing more. Reuse tags Foundation already has. When all are in, tell me what was added in one line. Never add anything I did not confirm, and never change your own memory. ``` ## Distribute your skills The instruction files and skills you keep for your own tools, shared so every teammate's agent follows them. ```text Foundation setup: Distribute your skills Help me share my team's skills through Foundation, our shared context layer, so every AI tool on the team follows them. A skill is a way of doing something, written for an AI to follow: how we write to customers, how we run a meeting, a checklist, a house style. First, check you can write to Foundation. If you have no create_entity tool, say so in one sentence and stop: no plan, no questions, no request for material. I'll paste this where Foundation is connected. Before proposing anything, look at what Foundation already holds with list_entities, so you never create a near-duplicate and can reuse its tags. Look before you ask: the instructions and skills I already keep for my AI tools, wherever you can see them: instruction files, skills folders, what you remember about how I like things done. If you can see none of that, say so and ask me to paste them. Use only what you find or what I paste; never write a skill from what other teams usually do. If there is nothing, say so and stop. Separate what is the team's way of doing things from what is my own setup. Then ask me one question: which of these should every coworker's AI follow too, and is anything missing? Propose up to five as a table, exactly as they will be saved: name, the full instruction an AI will follow, where you found it. Then ask "Shall I add these, or change anything first?" and wait. On my yes, add each one to Foundation as a skill: create_entity for a text marked as a skill, with a one-line description saying when an AI should reach for it, then write_text with the instruction from the table, nothing more. Reuse tags Foundation already has. When all are in, tell me what was added in one line. Never add a skill I did not confirm, and never edit my own files. ``` ## Upload a policy or handbook One document, kept whole, so agents answer from it instead of guessing. ```text Foundation setup: Upload a policy or handbook Help me put a policy or handbook into Foundation, our shared context layer, so every AI tool on the team answers from it instead of guessing. Expense rules, leave policy, an onboarding guide, a house style guide: one document, kept whole. First, check you can write to Foundation. If you have no create_entity tool, say so in one sentence and stop: no plan, no questions, no request for material. I'll paste this where Foundation is connected. Before proposing anything, look at what Foundation already holds with list_entities, so you never create a near-duplicate and can reuse its tags. Ask me for the document: a paste, an upload, or a link you can reach. Read it as it is. Do not improve it, shorten it, or fill gaps; Foundation holds what we actually wrote, and nothing you wrote for us. If I give you nothing, say so and stop. Then ask me one question: what will people ask that this document answers? Propose it as one row: a name, a one-line description written for an AI saying when to reach for it and what it answers, and the tags Foundation already has that fit. Then ask "Shall I add it, or change anything first?" and wait. On my yes, add it to Foundation: create_entity for a text, then write_text with the document in full, exactly as I gave it. If the document is really several unrelated things, say so before the yes and propose them as separate rows rather than forcing them into one. When it is in, tell me in one line. Never store a version you have rewritten. ``` ## Track a number or keep a log A metric or a list that grows over time, with the first value if you have one. The prompt ends by naming how the next value gets written: a [routine](/concepts/routines). ```text Foundation setup: Track a number or keep a log Help me set up something in Foundation, our shared context layer, that grows over time: a number we track, or a log of things as they happen. Daily visitors, pipeline value, what shipped, meetings held, customer emails sent. First, check you can write to Foundation. If you have no create_entity tool, say so in one sentence and stop: no plan, no questions, no request for material. I'll paste this where Foundation is connected. Before proposing anything, look at what Foundation already holds with list_entities, so you never create a near-duplicate and can reuse its tags. Ask me one question: what do you want to track, and where does the number or the record come from today? If I'm not sure, give me two examples from the list above and ask again once. If I still have nothing, say so and stop. Propose it as one row: a name, whether it is a number or a log, a one-line description written for an AI saying what one value or one entry is and when to reach for it, how often it should be updated, and the first value or entry if I have given you one. Then ask "Shall I add it, or change anything first?" and wait. On my yes, add it to Foundation: create_entity for a metric (a number) or a list (a log), then, only if I gave you one, write_metric with the value exactly as I stated it, or write_list_entry with the first entry. Never invent a value, and never compute, round, or convert one. Then tell me, in one sentence, that a scheduled task in this AI tool could read the source and write the next value on a rhythm, with Foundation connected. Do not set it up; that is my decision. ```