Andy University coding agents for real beginners

Guide Note 1

Your Personal Rules For The Agent

Teach the agent how you actually like to work instead of accepting a generic default helper.

Open the short practice promptBack To Start Here

Short Version

What this means if you are still new.

Start with three things: how much detail you want, what kind of output is most useful, and what annoys you when the agent answers.

Example

Sales example: “When I ask for outreach options, give me exactly three. Keep them short. Tell me which one you would send. Do not write a wall of theory before the actual draft.”

Open a short practice prompt Use this to feel the lesson without reading the full notes first.

Pasteable Prompt

Try this today

Help me write a short rule sheet for how you should work with me.

Include:
- how concise you should be
- how I want options presented
- what you should avoid

Make every line specific enough that I would notice if you broke it.

This is the shortest possible action that lets you feel the lesson in practice.

Longer Notes

Open the full guide notes only when you want more depth.

This keeps Andy's original longer write-up available without forcing all of it onto a first-time visitor.

Open the full notes for guide note 1 Andy calls this a constitution. This is the longer original version.

Your AI doesn't know you.

Not a problem with the AI. It's a problem with the setup. The AI has been trained to be helpful to everyone, which means it's optimized for no one. It gives the answer that works for the median person in the median context with median stakes.

The constitution is how you fix this.

What It Is (and Isn't)

A system prompt is static. Write it once, paste it at the start of sessions or leave it in custom instructions. It tells the AI to be concise, professional, whatever. It makes sessions slightly less generic.

A constitution is a live document. It tells the AI who you specifically are, how you make decisions, what your vocabulary means, and what it should never do in your sessions. It gets better as you develop it. The AI running on a year-old constitution is a qualitatively different collaborator than the one running on your first version.

The standard: specific enough to be wrong. Every line in your constitution should be something the AI could clearly violate. If you can't imagine a violation, the line isn't doing anything.

The Five Components

Identity and Role

Not biography — operational context. What are you working on, where do you have leverage, what constraints bind you?

Weak version: "I'm a product manager."

Strong version: "I'm a PM at a startup, own the roadmap, no direct engineering authority. My leverage is defining the problem space before engineering gets involved — once they're scoping, options narrow fast. I need help thinking through the problem, not jumping to solutions."

The strong version tells the AI where you have leverage and what kind of thinking you actually need.

Decision Logic

How you make decisions. Revealed preference, not stated preference.

Weak version: "I value quality and speed."

Strong version: "When choosing between options, I weight reversibility first — a reversible choice I can make fast. I also weight what I can actually execute given the team I have. When I ask 'which option,' give me the most reversible one unless I've said the decision is time-critical."

Test: if you gave your decision logic section to someone who doesn't know your work, would they make decisions similar to yours?

Communication Preferences

When to be concise, when to be thorough. What you do with different output formats.

Example: "Bullets for options, paragraphs for analysis. Mid-session, give me the minimum to keep going — save the background for when I ask. If I ask 'why,' give me the root cause, not the proximate one."

Vocabulary

Terms that have specific meaning in your context. Every time you use a term in a session and have to explain it, add it to this section. Over time this becomes one of the highest-leverage parts.

What to Avoid

Negative constraints are often the most specific. Examples:

  • "Don't preamble my own question back to me. Start with the answer."
  • "Don't hedge every statement with 'it depends' — give your best judgment."
  • "Don't suggest I 'consult a professional' for domain questions in my own field."

The Living Document Principle

The constitution isn't written once. After every session, ask one question: "Is there anything I had to explain twice?" If yes, add it to the constitution.

The update is the work. A constitution you wrote six months ago and haven't touched is a constitution that's been drifting from how you actually work for six months.

Practice

Write your version 1 constitution. Minimum 500 words. No polish — just get the five sections down.

Then load it into your next AI session and notice:

  • One moment when the AI did what you wanted without explanation
  • One moment when it still didn't

The first is evidence it's working. The second is your next update.

Completion criteria: A constitution file exists. One session run with it loaded. One update made based on that session.

What You've Built

A starting point. Version 1 will be too generic in some places and too specific in others. That's right. The version that's actually good is version 10, reached through 9 sessions of noticing what's wrong and updating.

The trajectory, not the document, is the thing.