Build an AI advisor for career decisions
A folder with four text files is all it takes to give an AI enough context to be genuinely useful when you face decisions that actually matter.
The appeal of an AI advisor is obvious. The execution usually falls apart. Most people either dump a wall of personal context into a system prompt and wonder why the advice feels generic, or they start fresh every session and get outputs that could apply to anyone. Neither approach is what we mean here.
To build an AI advisor that gives useful, personalized guidance on life and career decisions, you need four structured files: one that defines how the AI should advise you, one with your goals and context, one where it accumulates patterns over time, and one that sets evaluation criteria before it responds. Together, these turn a general-purpose model into something that actually knows you.
The newsletter writer Peter Yang documented his own version of this setup in detail, and we think his four-file architecture is the right frame. But our angle is different from his. Yang built this to navigate a personal career pivot. Our readers are B2B founders and senior leaders who need sharper judgment on business decisions: when to hire, when to say no to a deal, whether a market signal is real or noise. The same structure applies, but the content inside each file looks completely different when your decisions carry revenue consequences.
Why most AI advice is useless
When you open a chat with any AI model cold, it has no idea who you are. It knows you asked a question. It answers based on the average of everything it was trained on, which means the advice is calibrated for no one in particular.
The problem is not that the models are bad. It is that they are missing the context that makes advice actionable. A good human advisor does not just answer the question you asked. They think about your history, your constraints, your psychology, your stated goals, and the gap between what you say you want and what you actually do. That is what the four-file setup replicates.
For founders and leaders, the stakes are real. A decision made with incomplete context costs pipeline, hires, or momentum. Getting AI to think alongside you on those decisions is worth the setup time.
The four files
skill.md: how you want to be advised
This file tells the AI how to behave as an advisor. Not who you are. How it should respond to you.
The distinction matters. A lot of people conflate persona context with behavioral context. Your skill.md should answer questions like: Do you want the AI to push back, or surface options? Do you want it to ask clarifying questions before it advises, or give a direct take first? Should it flag when your reasoning seems inconsistent with your stated principles? Should it tell you when a decision looks emotionally driven?
For B2B founders, a useful skill.md might include:
- Give the business consequence first, then the reasoning behind it.
- Flag when I am anchoring to sunk costs rather than forward value.
- If I present only one option, ask what I am ruling out and why.
- When I describe a team problem, distinguish between a systems issue and a performance issue before advising.
- Do not validate a decision I have already made. Tell me what I might be missing.
This file stays relatively stable. You update it when you notice the AI's advice style is not working, not every session.
plan.md: your actual context
This is the file that does most of the work. It is your living document of goals, constraints, priorities, and the information the AI needs to give advice that fits your actual situation.
For a B2B founder or senior leader, plan.md might cover:
Current business state. Revenue run rate, pipeline health, team size, the two or three things that are working, and the one or two that are not. This is not a pitch deck. It is an honest operational snapshot.
Goals with a time horizon. Not "grow the company." Something like: close the year at $X ARR, make a hire decision on VP Sales by Q3, get the product to a state where we can reduce founder-led sales. Specific and time-boxed.
Principles you have already decided. These are the things you do not want to re-litigate every time. The AI should treat them as constraints, not suggestions. Examples: we do not take deals where the buyer has no internal champion, we do not hire fast and fire slow, we do not trade equity for non-cash deals.
Energy and capacity. What is consuming you right now. Where you are stretched. What decisions you are avoiding. This sounds personal, but it is operationally relevant. Founders who are spread thin make different kinds of mistakes than founders who have head space.
Key relationships and dependencies. Co-founder dynamics, board relationships, a critical hire that is uncertain. Not gossip, just the human variables that affect decisions.
You update plan.md when something materially changes: a big win, a team change, a strategic shift, a quarter that went sideways. Think of it as a quarterly refresh with ad hoc updates when something significant happens.
learnings.md: accumulated patterns
This is the file that makes the advisor better over time. After a significant conversation or decision, you add what you learned to this file.
The format is simple: date, the decision or situation, and the pattern it revealed. Some examples of what this looks like for a B2B operator:
- June 2025: Took on a client outside our core segment because of revenue pressure. They churned in four months. Pattern: desperation deals cost more than they pay.
- September 2025: Delayed a hard feedback conversation with a senior hire for six weeks. Problem got worse. Pattern: I am conflict-averse with people I respect.
- November 2025: Followed a competitor's pricing move without validating that our buyers cared. Pattern: I react to competitive signals faster than I should.
Over time, this file becomes a pattern library of your actual decision-making. The AI can use it to say: "You have flagged conflict avoidance three times in learnings.md. Is that what is happening here?"
That kind of callback is not possible in a stateless session. It is the thing that separates a real advisor relationship from a one-off consultation.
eval.md: the pre-advice checklist
Before the AI gives you a substantive response, eval.md defines what it should check. Think of it as the AI's internal rubric before it speaks.
For business decisions, a useful eval.md might include:
- Have I identified what kind of decision this is (reversible vs. irreversible, time-sensitive vs. deferrable)?
- Am I giving advice based on the user's stated principles, or am I defaulting to generic best practices?
- Have I surfaced the second-order consequence, not just the immediate one?
- Am I recommending action when the right answer might be to wait?
- Is there a version of this decision the user has already faced? (Check learnings.md.)
This file keeps the AI from giving fast, confident answers when the situation calls for slower thinking. It is the equivalent of a surgeon's pre-op checklist: not a sign of doubt, just discipline.
How to set it up in practice
You do not need special software. This works in Claude (with Projects), in ChatGPT (with custom GPTs), or in any interface that lets you attach persistent context.
The simplest version: create a folder on your computer called /advisor. Put the four files in it. When you start a session, paste or attach the relevant files depending on the decision at hand. For a strategic business question, you probably need plan.md and eval.md. For a leadership or team decision, you might need all four.
The setup takes about two hours the first time. Writing skill.md takes thirty minutes if you actually think about how you want to be advised, not how you want to be flattered. Writing plan.md takes longer because most founders have never written down their constraints and principles in one place. That exercise alone is worth doing regardless of the AI use case.
Peter Yang's tutorial references running this through Claude's interface with a /advisor command. That workflow works well if you use Claude regularly. The structure is the same regardless of the model.
One note on models: for decisions with real stakes, run the conversation with a frontier model (Claude Opus, GPT-4o, Gemini Ultra). The difference in reasoning quality on ambiguous, multi-variable problems is material. This is not the place to use a smaller model to save tokens.
What this is not
A personal AI advisor built this way is a thinking tool, not an oracle. It is useful for surfacing what you are not seeing, stress-testing your reasoning, and catching patterns you repeat. It does not know your industry better than you do, and it cannot substitute for talking to people who have done what you are trying to do.
The founders we work with who get the most out of setups like this treat the AI as a sparring partner, not a decision-maker. They use it to pressure-test a position before they bring it to a board, a co-founder, or a key hire. That is the right use case.
Used that way, this is one of the more useful things you can spend two hours building. Most productivity tools ask for your time in exchange for marginal gains. A well-built AI advisor gives you back the quality of thinking you usually only get from a trusted mentor, available whenever you actually need it.
For more on how we think about using AI in business contexts, see our piece on how we use an AI marketing assistant and our breakdown of AI strategy and the trust problem.
Frequently asked
The core setup is four files: a skill file that defines how the AI should advise you, a plan file with your goals and constraints, a learnings file where patterns accumulate over time, and an eval file that sets criteria before the AI responds. Load these into any AI interface that supports persistent context, like Claude Projects or a custom GPT. The personalization comes from the quality of what you put in plan.md and how consistently you update learnings.md after significant decisions.
