Give it your context
The suggestion that is right in general and wrong here
Ask an AI for ideas to grow your business and you will get a competent list. Partnerships, referral programme, tiered pricing, content, a webinar.
None of it is wrong. Most of it is useless, because it does not know that you have no marketing headcount, that your largest client's contract forbids naming them, or that you tried a webinar in March and three people came.
The build told it who you are. This lesson tells it where it is standing.
Two ways to do this
You can interview your way to these files, or fill them in like a form. The interview is faster for most people. If you would rather have the form, ask for that instead: "Write me a COMPANY_CONTEXT.md template with headings and a one-line hint under each, and I will fill it in myself." Either route ends up in the same place, and nothing gets downloaded either way.
Same as last time, you can just ask:
Interview me about the business now. What it does, who buys, and what constrains what we can actually do. One question at a time. When you have enough, write COMPANY_CONTEXT.md and GOALS.md into my AI OS, and show me what you wrote.
COMPANY_CONTEXT.md
Ten minutes, and one section matters more than the rest.
The straightforward parts
What the business does, in one sentence. The way you would say it to someone at a barbecue, not the way the website says it
How it makes money. Be specific about which products actually produce revenue, and which ones people assume are big but are not
Who buys. And who does not, though people assume they do
The part that earns its keep
Constraints. This is the section that turns a plausible suggestion into a useful one, and it is the one everybody skips.
- Regulatory. Anything governing what you can claim, store or say. If you are
in health, finance, law, education or anything licensed, write it down properly
- Budget. The real number, not the aspiration
- People. What there is no capacity to do, however good the idea
- Technical. Systems that cannot be changed, integrations that do not exist
- Contractual. Exclusivity, non-competes, clients you cannot name publicly
- Political. Things that are perfectly possible and would not survive the room
That last one is worth writing down even though it feels odd to put in a file. A lot of bad advice is technically sound and politically dead on arrival.
Your words
Ask for a small table at the end for terms your industry uses in a way an outsider would get wrong. Fill in three or four. It stops the AI writing confidently in a register that marks you as an outsider in your own market.
GOALS.md
Five minutes. Three goals you mean beats ten you listed.
For each one:
- Done looks like. A number, a date, or a thing that exists. Not "improved"
- By when
- Currently. Where it stands today, so progress is measurable
- In the way. What is blocking it
Then the section most people skip, which is the most useful one here: what you are deliberately not doing. Things worth doing that you have chosen not to do now, and roughly when you would revisit.
Without that, the AI will keep proposing the thing you already considered and rejected, and you will keep having the same conversation.
The one instruction that makes this pay off
Once both files exist, start asking for this:
A good response to that is sometimes: this does not serve any of your goals, and here is the goal it would crowd out.
That is the moment the thing stops being a writing assistant.
KNOWLEDGE, and one document
The KNOWLEDGE folder is for things that are true about your world and that you would otherwise re-explain or go hunting for. Policies, pricing, brand guidelines, research, a past piece of work worth reusing.
Add one real document now. Whatever you have explained to an AI more than twice.
Then put these three lines at the top of it:
> **Source:** [where this came from]
> **Accurate as at:** [date]
> **Confidence:** [high, or "rough, check before quoting"]Three lines, and they are the difference between a useful folder and a confidently wrong one. An AI cannot tell that your 2024 rate card has been superseded unless you say so. It will quote it with exactly the same confidence as everything else.
A stale knowledge folder is worse than an empty one. Once a quarter, ask it: look through KNOWLEDGE, which documents are more than a year old and which ones contradict each other.
Stuck anywhere in this lesson? Reply to any Build Notes email and tell me where. I would rather fix the lesson than have you quietly give up.