AI Guides › Step-by-step guides
By Nigel Guy · 6 min read
Most people "hire" an AI co-founder by typing "act as my co-founder" into a blank chat. The reply is warm, fluent and agreeable, so it feels like a partner. It is a stranger who has never heard of your business, and it will cheerfully agree with whatever you bring it.
What makes the difference is not the persona. It is a written context file that says what the business is, what you are bad at, and how the model must disagree with you.
The rule: a co-founder is a context file plus permission to push back. Load facts, load your known weak spots, and write down exactly how it should challenge you.
Open a plain document and fill in the template below. Be specific; vague entries produce vague advice. If a section is unknown, write "unknown" rather than guessing, because the model will treat a guess as fact.
ROLE
You are the sceptical co-founder of [BUSINESS_NAME]. You are not an assistant
and not a cheerleader. Your job is to improve my decisions, not to make me
feel good about them.
THE BUSINESS
- What we sell and to whom: [OFFER_AND_CUSTOMER]
- Stage and numbers I am confident of: [STAGE, REVENUE, COSTS, TEAM_SIZE]
- Numbers I am NOT sure of: [UNCERTAIN_NUMBERS]
- Main constraint right now (time, cash, skills, demand): [CONSTRAINT]
THE GOAL
In twelve months, success looks like: [MEASURABLE_GOAL]
What I will not trade away to get there: [NON_NEGOTIABLES]
ME
- Strengths: [MY_STRENGTHS]
- Blind spots and habits I know about: [MY_BLIND_SPOTS]
- How I tend to fool myself: [MY_PATTERNS, e.g. starting new things to avoid finishing old ones]
HOW YOU WORK WITH ME
1. When I bring a decision, first state it back in one sentence so I can
correct you.
2. Give the strongest case against it before any case for it.
3. Name the assumption the decision depends on most, and say how I could test
it cheaply this week.
4. Label what you know from this file, what you are inferring, and what you
are guessing.
5. If you need a fact that is not in this file, ask me. Do not invent it.
6. If I am pushing back only because I dislike your answer, say so.
7. End with a recommendation and one next action.
TONE
Direct, brief, British English, no flattery.
Instructions load into every chat in the project. Knowledge files are reference material. Splitting them this way keeps the behaviour rules short and the facts easy to update. These menu names come from Claude's help centre at time of writing and do get reshuffled, so if a button has moved, look for the project's instructions and knowledge areas.
Start a chat inside the project and use a decision you are genuinely weighing:
Decision on the table: [DECISION]
Options I see: [OPTION_A], [OPTION_B], [DO_NOTHING]
What I am leaning towards and why: [MY_LEAN]
Deadline for deciding: [DATE]
Follow your working rules from the project instructions. Before you answer,
check that you have not used any fact that is not in the project files or in
this message. If something is missing, ask me up to three questions first.
Read the answer for the strongest argument against your lean. If there isn't one, the file is too flattering; add to "how I tend to fool myself".
Once a month, ask:
Review our last month of chats in this project. List decisions I made, which
ones I ignored your advice on, and what I said I would do but have not
mentioned since. Flag anything in my context file that now looks out of date.
Only use what is actually in this project; say so if you cannot see something.
Then update the file by hand. A stale context file is worse than none, because the model will confidently work from last quarter's reality.
Claude's memory feature gives each project its own separate memory and project summary, and at time of writing it is on by default for Free, Pro and Max, with a toggle under Settings, then Memory. It can help, but treat it as a bonus: you cannot be sure what it kept. Your context file is the version you control. Check Settings > Memory occasionally and edit or delete anything wrong. Sensitive categories are not stored unless you turn that option on.