Anyone who has lived with a cat knows the rule: you can’t teach instinct. Either the animal has it or it doesn’t. What you can do is set up the yard, so the instinct has something useful to chase.
AI tools are the opposite case. They have plenty of raw ability and no instinct about your business at all. They don’t know who your best client is, what you refuse to say, or why your proposals open with a question instead of a bio. Left on defaults, they produce average work in an average voice. The good news: the part you can’t hand over (your judgment) is small, and the part you can train (everything around it) is large.
Instinct vs. Procedure: Know Which Is Which
Sort your work into two piles.
- Instinct: deciding what to build, who to say no to, what to charge, what your brand stands for, when a client is about to churn. This comes from experience and skin in the game.
- Procedure: turning a rough note into a clean client update, formatting a proposal, summarizing a call, drafting the first pass of a blog post, tidying a spreadsheet.
Training an AI tool is the job of writing your procedure down so well that a smart stranger could follow it. If you can’t explain how you do something, the tool can’t copy it. That discovery is half the value.
The Five Training Assets
1. A style guide (one page, not a manual)
Write down how you sound and how you don’t. Cover:
- Who you write for, and what they already know.
- Sentence habits: short, blunt, contractions, a dry joke now and then.
- Words and moves you ban (the “in today’s fast-paced world” openers, hype adjectives, fake urgency).
- Formatting rules: headings, bullets, how long a typical email runs.
Explain the reason behind each rule, not just the rule. Anthropic’s own prompting guidance notes that giving the reasoning behind an instruction helps the model generalize to cases you didn’t list. “Never use jargon” is weaker than “my readers are shop owners who bounce at jargon.”
2. An example library
Examples beat adjectives. “Write in a friendly, professional tone” means nothing; three of your actual emails mean a lot. Both Anthropic’s and OpenAI’s prompt guidance recommend including examples, and Anthropic’s suggests roughly three to five that are relevant to the task and varied enough that the model doesn’t latch onto an accidental pattern.
Build the library by saving your best real work as you go: the proposal that won, the email that got a reply, the onboarding message clients thank you for. Strip out client names and anything confidential. Label each one with what makes it good. Keep a few “bad” examples too, marked as bad.
3. Saved prompts (your templates)
If you’ve typed the same request three times, it’s a template. A good saved prompt has five parts:
- Role and goal: what this task is for.
- Inputs: what you’ll paste in (call notes, a brief, raw numbers).
- Style guide and two or three examples attached or referenced.
- Output format: length, headings, what to flag.
- An escape hatch: “If something is missing or unclear, ask me or mark it [CHECK]. Don’t guess.”
Store them in one place you control (a notes app, a document, a folder of text files), not buried in a chat history. Name them like files: client-update-weekly, proposal-first-draft, call-summary. Most AI tools now offer some form of saved instructions or project workspace, but tools change; your library in plain text moves with you.
4. A feedback loop
This is where most solo operators stop too soon. Every time you edit an AI draft, you’re producing training data. Use it.
- When you fix the same thing twice, write the fix into the style guide or the prompt.
- Keep a running “corrections” note: too salesy in the opener, invents statistics, forgets to mention the deadline.
- Once a month, spend 20 minutes folding that list back into your assets, and swap in a newer, better example.
It’s a habit like any other, so give it a recurring slot. (See developing great habits for how to make it stick.)
5. A tiny test set
Pick three or four real inputs you’ve handled before, with outputs you were happy with. Whenever you change a prompt, or the tool updates and the results suddenly feel off, run the same inputs and compare. OpenAI’s guidance recommends exactly this idea (build evals and track results as you iterate or change models), and for a one-person shop it can be as simple as a folder with four test cases. Without it, you’ll notice quality slipping only when a client does.
A Worked Example (Hypothetical)
Say you run a one-person brand consultancy and you send each client a weekly update. Today it takes you 40 minutes of rewriting scattered notes. Here’s the trained version:
- You paste your raw notes: bullets, half sentences, a screenshot’s worth of numbers typed out.
- Your saved prompt supplies your voice guide, a template (wins, blockers, next steps, one decision needed) and two past updates the client liked.
- The AI drafts the update and lists anything it couldn’t confirm from your notes.
- You check every number against the source, fix anything that sounds like not-you, and send.
- If you made the same edit again, you update the prompt.
The human decides what matters and what’s true. The AI handles shape and polish. Nothing goes out unreviewed.
What a Solo Founder Does With This in the Real World
Pieter Levels, who has built Nomad List, Remote OK and other products on his own since his 2014 “12 startups in 12 months” challenge, wrote in mid-2026 that he now does almost all his coding with an AI coding agent working directly on his server. Two things in his account are worth stealing. First, he keeps proper backups, so a mistake is cheap to undo. Second, he says outright that a larger company should use a staging server so AI changes don’t touch production. That’s a trained, bounded setup: the tool does the typing, the human owns the stakes. What he decides to build, and whether it works, stays with him.
What to Never Delegate
Draw the line in ink:
- Positioning and pricing. The tool can model options. You choose.
- Promises. Deadlines, guarantees, scope and anything contractual.
- Hard conversations. Apologies, firing a client, a late-payment standoff. Draft help is fine; the voice and timing are yours.
- Confidential material you haven’t cleared. Check a tool’s data-handling terms before you paste in a client’s private information, and see stopping data leaks.
- The final read. Anything with your name on it gets a human review. More on that in Get It Right, Not Just Fast.
Not sure what’s safe to hand off? The delegation self-quiz works just as well when the assistant is software.
The Takeaway
You’re not going to teach a tool your instincts. Teach it your procedures: one-page style guide, a handful of real examples, a few saved prompts, a correction log, and a four-case test set. Start with the one task you repeat most this week. Build the pieces for that, run it for a month, and let the feedback loop do its work. Want a head start on the rest of the setup? Grab the free “One-Person Business AI Stack” checklist.
General information, not professional advice. Tools and their data policies change; check current terms before sharing client information.
Sources
- Anthropic, “Prompting best practices” (guidance on examples, context and role): platform.claude.com
- OpenAI, “Prompt engineering” guide (few-shot examples, evals, pinning model versions): developers.openai.com
- Pieter Levels, “I’m Launching 12 Startups in 12 Months” (March 2014): levels.io
- Pieter Levels, “I’ve been coding almost solely on my VPS with Claude Code for almost a year now” (June 2026): levels.io