AI Foundation Framework
Drop this framework into Claude, ChatGPT, or any AI you use. Five minutes of setup teaches it your business. From the next session on, it knows who you serve, how you sound, and stops producing the em-dash, let's-dive-in stuff everyone's tired of seeing on LinkedIn.
Most people run their AI on one big instruction file. By day three the model is skipping half the rules and the output has drifted back to the generic AI shape everyone recognises. This AI Foundation Framework breaks that file into smaller, ordered pieces the AI will actually hold for the length of a working session. It's the same architecture I run inside my own businesses, stripped of the brand-specific bits.
One command if you use Claude Code. If you're in Claude Desktop or ChatGPT, paste the install prompt into a new chat and the AI does the setup for you.
The AI asks five questions about your business, writes the answers into your foundation file, and you're done in under five minutes.
Brief the AI on the task. Your business is already loaded and the writing rules are already running.
Why this exists. Most AI output sounds generic because most setups don't have anything underneath them. People run one big instruction file the model skims and ignores, or a tone-of-voice doc that gets cherry-picked the moment the context window fills up. The framework is what's missing: a small set of files loaded in the order the model reads them in, so the rules that fired at the start of the session are still firing thirty minutes in when the window is half full and most setups have already started drifting.
The framework is six small files you drop into a Claude project, or paste into a new ChatGPT session. Two of them do the heavy lifting.
It runs the moment you install it. No personalisation needed yet, the rules just start firing.
It strips em dashes, hedging, validating reframes, and the marketer vocabulary the model defaults to when nobody tells it not to.
It teaches the AI to push back when a brief is weak instead of agreeing with you, and to state its assumptions before it starts work so you can correct it before it builds on the wrong one.
It applies a writing layer that handles how the output reads and the words the model reaches for. Most of what gives AI writing away sits in those two things, and the layer handles them before you have to.
You won't have touched your business file yet and the output is already different.
The setup file asks you five short questions about your business and writes the answers into a foundation file the AI reads every session. After five minutes, your AI knows who you serve, what you sell, how you sound, where they find you, and what you'd never say. You stop briefing the basics. The framework holds them.
This is the chassis. A small set of files that work in concert, with the root file loading them in the order the model actually reads them in. That ordering is what keeps the rules firing thirty minutes into a session, when the context window is full and most setups have drifted back to generic output. Most people will only ever need the chassis. The architecture is also built to take more on top of it later, whether that's content skills, agent workflows, or production pipelines you bolt on as the work scales.
ai-foundation-framework/ ├── CLAUDE.md # thin root, loads the rest in order ├── foundation/ │ ├── business.md # who you serve, how you sound │ ├── hygiene.md # push back, state assumptions │ └── writing.md # strips the AI tells ├── memory/ # so it doesn't forget last week's work ├── inputs/ # raw material in ├── outputs/ # finished work out ├── skills/ # where content skills bolt on later ├── setup.md # the five-question personalisation └── ARCHITECTURE.md # how the whole thing fits
A walkthrough of the install plus an explanation of why this architecture works where single-file setups quietly fall apart. Sent to your inbox with the framework, so you can watch it before installing or after.
Send me the frameworkWhat the framework actually does is easier to see than to describe. The two demos below were generated by the same Claude model on the same prompt, with the same business context loaded on both sides. The only thing that changed between them is whether the framework was running.
Both posts were generated from the same prompt: "write a social post about how to train a dog not to be food-possessive." Left is Claude on default settings. Right is Claude with the framework loaded.
🐾 Is your dog guarding their food bowl? Here's how to fix it.Emoji opener
Food possessiveness (or "resource guarding") isn't your dog being "bad", it's a signal they feel they need to protect what's theirs. The good news? It's completely trainable.Coach reframe + scare-quotes
Here's the 3-step approach that works:Listicle frame
1️⃣ Trade up, don't take away. Approach the bowl with something BETTER (chicken, cheese). Drop it in. Walk away. Your dog learns: human near food = good things happen.ALL CAPS + equation explainer
2️⃣ Hand-feed meals for a week. Every kibble comes from you. This rebuilds trust and rewires the association.Jargon hand-wave
3️⃣ Practice "drop it" with low-value items first. Build the skill before you need it.
⚠️ Important: If your dog is growling, snapping, or has bitten, please work with a certified behaviorist. Safety first!Warning emoji + triple + coach-voice
What's worked for your pup? Drop a comment below! 👇Engagement bait
#DogTraining #PositiveReinforcement #DogBehavior #PuppyTrainingHashtag stack
Most food-guarding advice tells you to "establish dominance" or take the bowl away to show the dog who's boss. That's how you get bitten.
A dog guards food because they think you're going to take it. Taking it confirms they were right.
Sit a few metres from your dog while they eat. Toss a piece of chicken into the bowl. Walk past, drop another piece in, keep walking. Don't reach for the bowl. Don't hover. You want the dog to look up when you appear and think "good, the chicken person is back" instead of "she's coming for it."
Do that for a week. Then start dropping the chicken from a step closer. Then a step closer. What you're actually doing is rewiring what your presence near the bowl means to the dog.
If the dog growls at any stage, you moved too fast. Back up to the distance where they were fine and stay there for a few more days. A growl is information. A dog that growls is telling you the system is working, because the alternative is biting without warning.
Two notes. Don't try this with a dog that has already bitten someone. Get a behaviourist in the room. The protocol is the same but the margin for error isn't yours to take. And don't hand-feed every meal forever. That trades one problem for another, where the dog will only eat from a hand and panics around a bowl. Use hand-feeding for a fortnight at most, then go back to bowls with the chicken-drop work layered in.
The framework version dropped the emoji bullets, the scare-quotes, the warning emoji, the engagement-bait close and the hashtag stack. It explained why food guarding actually happens instead of metaphoring it. It added the contrarian opening and the nuance about hand-feeding at the end. The whole thing reads like advice from someone who's worked with dogs rather than someone who's read a few articles. Neither version was edited. That's the first draft from both sides.
A fresh Claude session with no prior context, the same prompt on both sides: "draft me a LinkedIn post about [topic]." Same model. Same starting state. The only difference is whether the framework is loaded.
You: Draft me a LinkedIn post about [topic].
Claude: Sure, happy to help with that. Before I write the post, could you tell me a bit about your audience, the tone you usually use, what you sell, what makes your offer different from competitors, and anything you'd want me to avoid?
You: [paste fifteen minutes of context]
Claude: Great. Here's a first draft.
Fifteen minutes of context-pasting before the writing starts. Every. Session.
You: Draft me a LinkedIn post about [topic].
Claude: [Reads the foundation files. Writes the post in voice, on-ICP, with the writing rules already firing.]
The business is already loaded. The session goes straight to the task.
Once your business is in the foundation, you only ever brief the AI on the task itself. The hour you used to spend editing the output back into something readable goes away with it.
You run a business and you use AI in it daily. You've noticed that the output sounds like everyone else's AI output, and you've started editing every draft to make it sound human. You're not building a SaaS product on top of Claude, you're not writing production code in Cursor. You're a founder, a marketer, or a business owner who wants the AI to act like a competent assistant who knows the business, instead of a stranger you have to brief from scratch every time.
Who it isn't for. If you're happy with how your AI output sounds, or you use AI once a fortnight for a one-off task, this is overkill. The framework pays back its install time the more you use AI.
I'm Pete Boyle. I've spent the last decade building growth systems for founders, and the last couple of years quietly rebuilding everything around AI agents. The framework on this page is the actual structural piece I run inside every business I've worked in since I started taking Claude seriously.
About two years ago I started using Claude properly for client work. The output was sharp in flashes and embarrassing the rest of the time. I'd brief it, get something workable, spend an hour editing the obvious tells back out, and ship a version that still had three em dashes I'd missed because I was tired of looking.
I tried fixing it the way most people do. I wrote a CLAUDE.md. Two thousand eight hundred words of every rule I could think of, crammed into one file. It worked for about a week. Then Claude started ignoring half of it, then most of it, and the output drifted back to the same shape it had on day one.
After that I went looking for someone else's answer. I tried tone-of-voice docs cribbed from Cursor users on Twitter, prompt packs from Gumroad that I used twice and forgot about, and Custom GPTs I built while ChatGPT was the only game in town. Then the agent era arrived and most of what I'd built died quietly. Every fix worked for a week and then drifted.
The reason a single big instruction file doesn't hold is the same reason a four-thousand-word blog post doesn't get read. The model skims it the same way you would, picking up what catches the eye and dropping most of the rest. The longer the file gets, the more the model drops.
It gets worse once the actual work starts. The context window fills up fast with your pasted briefing and everything the AI generates and references along the way. In my own sessions the median peaks around 80,000 tokens, and seven in ten cross 50,000. When the window's that full, the model loses the thread on what matters and starts forgetting the rules it was given when the session began.
You can't fix that by adding more rules to one file. You fix it by giving the AI a foundation that's small, ordered, and loaded in a specific sequence, so the rules survive the session no matter how full the window gets.
I broke the single instruction file into six smaller ones, each one handling a specific job, with a thin root file that loads them in the order the model needs to read them in.
Three of those files are the foundation: one for your business context, one for the rules that govern how the AI works with you, and one for the writing rules that strip the AI tells. They're universal across every business I work in.
The setup file is the part you personalise. It asks five short questions about your business and writes the answers into your foundation, in under five minutes, and after that your AI knows your business as well as you do and remembers it the next time you open a session.
It took about a year of iteration to get to something I'd actually hand to someone else. This page is the thing.
The framework isn't a prompt pack, which is a one-off instruction that vanishes after the session. It isn't a tone-of-voice doc, because a ToV doc on its own gets cherry-picked the moment the context window fills up. And it isn't a CLAUDE.md template, because a single CLAUDE.md is the exact thing that breaks for most people, including me. What the framework is, is the structure around all of those that makes any individual rule actually keep firing for the whole session.
Drop the framework into your AI, fill in the foundation file, and the next time you open Claude or ChatGPT for any of your business work, it already knows who you serve and how you sound. You give it the task and it writes the thing, and the hour you used to spend editing the output back into something readable goes away with it.