Getting Started with AI, Seriously

Where Everyone Starts

You open a new chat. You type something like “write me a LinkedIn post about the leadership lesson I learned this week” and you get back something that is, honestly, not bad. It's structured. It hits the expected beats. It's just not you. It reads like a LinkedIn post written by someone who has read a lot of LinkedIn posts, which is exactly what it is.

So you try harder. You write a longer prompt. You include tone guidance: “be direct, not corporate.” You add constraints: “no bullet points, no inspirational platitudes.” You get a better result. You save that prompt somewhere. Maybe you iterate on it a few more times until you have something decent. You call it good.

This is progress. Specific prompts do produce better output than vague ones. That part is real. But it's also a ceiling, and most people hit it without realizing there was a floor above it they never reached.

The Problem with Every Session Starting from Zero

Here's what a good prompt cannot fix: the AI doesn't know you. Every session, you're reintroducing yourself to a blank slate. You explain your tone, your context, your industry, your audience, your preferences. Even if you've copy-pasted in a solid system prompt, you're still working with a model that has no memory of your actual writing, no sense of what you've said before, no awareness of the specific things you care about in your specific world.

The output reflects that. Generic AI copy has a texture to it that people recognize even when they can't name it. It uses the right words but not your words. It makes the right moves but not your moves. And when you need the AI to build something with current relevance, like a monthly content calendar that references what's actually happening in your industry right now, a prompt alone can't solve the sourcing problem. You either paste in a wall of context every single time, or you get output that's drawing on training data from a year ago.

The prompt-optimization loop has diminishing returns. At some point, you need a different architecture, not a better prompt.

The Unlock: Persistent Context

Claude Projects (and the equivalent custom GPTs in OpenAI's world, though I'll focus on Claude because it's what I actually use) solve this at the structural level. Instead of a chat session, you're building a project: a persistent workspace with its own system prompt, its own uploaded files, and its own context that carries across every conversation inside it.

This is not a minor quality-of-life improvement. It changes what AI can actually do for you.

The system prompt in a Claude Project isn't just instructions prepended to your chat. It's the permanent ground truth the model operates from every time you open that project. It knows your voice, your context, your constraints, your audience, before you type the first word. The uploaded files give it material to work from: examples of your writing, reference documents, source lists. It's the difference between hiring someone who read your brief once and hiring someone who has spent weeks embedded in your work.

I prefer Claude Projects over custom GPTs for a few reasons. The context window is larger, which matters when you're uploading real documents. The instruction-following is more reliable in my experience, especially for voice and tone constraints. And the file upload behavior is more predictable. Both tools will get you past the ceiling I described above. Claude just gets you further.

How to Build One That Actually Works

I'll use the content writing use case because it's concrete and most people can relate to it, whether you're publishing on LinkedIn, writing a newsletter, or producing thought-leadership content for your organization.

Step one is the system prompt, and you should not write it yourself.Most people sit down, try to describe their own voice, and produce something that sounds like a job description. “I am a [title] with [X] years of experience who writes about [topic] for an audience of [audience]. My tone is professional but approachable.” That's not a voice. That's a LinkedIn summary.

Instead, have the AI interview you. Open a blank Claude chat (not your project yet) and start with something like: “I want to build a detailed system prompt that captures my writing voice and context for a content project. Interview me to get what you need. Ask one question at a time.” Then actually answer the questions. What do you write about and why? What do you hate seeing in other people's writing? What's the most common mistake people make when writing about your area? Who are you trying to reach? What do you want them to feel or do after reading?

A good interview takes fifteen to twenty minutes. At the end, ask it to synthesize everything into a system prompt. Then edit that prompt until it feels accurate. This process produces something meaningfully better than what most people write themselves, because it surfaces things you know but wouldn't have thought to include.

Step two is uploading examples of your own work.Find five to ten pieces of content you're proud of. Posts where the voice landed. Writing that felt like you. Upload them to the project. The model will use them as stylistic reference. This is not about the model copying your sentences; it's about it developing a feel for how you construct an argument, where you put the weight, what you tend to leave out. The difference in output quality between a project with examples and one without is significant.

Step three is uploading reference material.This is the part most people miss entirely, and it's where the content calendar use case gets genuinely useful. If you want your monthly content calendar to draw on current events and real signal, you need to give the AI access to your sources. That means uploading a reference document that tells it: here are the subreddits I monitor, here are the newsletters I read, here are the industry feeds worth pulling from, here are the specific topics I want to stay current on. When you ask it to build next month's calendar, it now has a starting point that's grounded in your actual information diet, not just general vibes about your industry.

This document doesn't need to be elaborate. A simple list works. What matters is that it exists and that you keep it current. When a new source becomes relevant, add it. When something goes stale, remove it. The project is only as good as what you've given it.

What the Workflow Actually Looks Like

Once the project is set up, the workflow changes. You're not prompting from scratch; you're having a working session with a collaborator who already knows the context.

For a content calendar, you open the project and ask it to draft next month's calendar. It pulls from your reference sources, applies your voice guidelines from the system prompt, and builds something that actually looks like your content rather than a generic editorial template. You react to it, redirect it, tell it what's landing and what isn't. The iteration is faster because you're not correcting for missing context; you're just refining the output.

For individual posts, the same logic applies. “I want to write about the fact that most AI transformations fail not because of the technology but because of the org structure.” The project knows your voice, knows your audience, knows what you tend to do well and what to avoid. The first draft is closer. Not perfect, but closer.

The other thing that changes: you stop losing work. When you have a good conversation in a project, it's there. You can reference earlier drafts, earlier decisions, earlier instructions you gave. The project accumulates history in a way that a series of disconnected chat sessions never can.

This Applies Beyond Content

I used content writing as the example because it's relatable, but the same structure works anywhere you want AI to be a real collaborator rather than a sophisticated autocomplete.

Running customer support? Build a project loaded with your product documentation, your common-issue playbooks, your refund and escalation policies, and your house style for tone. Now a draft reply to a frustrated customer is grounded in how your product actually behaves and what you're actually allowed to offer, not a generic apology template. Paste in the ticket, get back something a human can send after a glance instead of a rewrite.

Want help on strategy documents? Build a project with your company context, your current priorities, your competitive landscape, your stakeholder constraints. The output reads like something produced by someone who has done the reading, because it has.

It doesn't have to be work, either. The same three inputs make a project out of almost anything you want to stay consistent at. A personal-finance project that knows your budget categories, your accounts, and the rules you're trying to hold yourself to, so “here's last month's spending” comes back as a read against your actual goals instead of generic advice. A nutrition project that knows your targets, your restrictions, and what you actually like to eat, so it plans meals you'll really make. Whatever the thing is that you keep doing and want done in a way that's actually yours, this is the structure.

The pattern is always the same: system prompt built through interview, examples of the output you want, reference material for the context it needs. Three inputs. Every project that actually works has all three.

Don't Start From Scratch — Take Mine

The interview is the hard part, and a blank page is where most people quit. So I'm giving you mine. The full discovery interview I use to build my own content engine, and the system prompt template it feeds into, are below. No gate, no email, no “comment KIT and I'll DM it to you.” People charge four figures for less than what's in these.

I've also adapted the same interview for two of the use cases above, so you're not left translating a LinkedIn prompt into a support or finance one yourself. These are starting points, not scripture. Paste one into a fresh chat, run the interview honestly, and edit the synthesis until it sounds like your situation instead of mine.

One thing the LinkedIn kit says that applies to all of them: the interview is only as good as the honesty you bring to it, and the output is a starting point you finish, not a post you publish verbatim. The engine gives you the idea. You bring the real story.

The Maintenance Mindset

One thing worth saying plainly: a project is not a one-time setup. It degrades if you ignore it.

Your voice changes. Your context changes. The sources you care about change. If you built your content project eight months ago and haven't touched it since, you're working with a snapshot of yourself from eight months ago. That's fine for stable contexts. It's a problem when your focus has shifted or your audience has evolved.

I treat projects like I treat documentation: periodically wrong, always worth fixing when you notice it. When an output feels off in a specific way, that's a signal. Usually it means the system prompt needs an update or a new example needs to be uploaded. The fix is usually five minutes. The payoff is that the next twenty sessions run better.

There's a version of this that people get wrong, where they spend so much time perfecting the project setup that they never actually use it. Don't do that. Build the first version in an hour, use it, update it when you see what's missing. A working project that's 70% right is worth more than a theoretical perfect one that you haven't built yet.

The Actual Gap Most People Are In

If you've been using AI for a while but still feel like it doesn't quite get you, this is almost certainly why. You're using a tool that resets every session and hoping that better prompts will close the gap. They won't, not fully.

The gap isn't your prompting. The gap is architecture. You're running a tool without giving it the structure that would let it actually know your work.

Claude Projects are free with a Claude account. The interview technique for generating a system prompt costs nothing but time. Uploading a few examples of your own writing takes twenty minutes. There is no technical barrier here. The only thing between most people and an AI that genuinely helps them is the decision to build something instead of just prompting.

Build the project. See what changes.