A structured framework to configure an AI assistant that speaks like you — no fluff, no warm-up time, no drift.
You're already using ChatGPT, Claude, or another AI to write your emails, posts, or proposals. But every single time, you waste precious minutes:
- Re-explaining your job, your clients, your industry,
- Fixing a tone that's too generic or too corporate,
- Deleting hollow filler phrases like "In today's fast-paced world…"
This framework is the answer.
💡 What is an AI assistant, for beginners? An AI assistant (like ChatGPT or Claude) is a software tool you can write to in plain language. You ask it a question or give it a task, and it generates a text response. Think of it as a very fast, very well-read writing partner — but one with no memory of you, no knowledge of your style, and a tendency to sound corporate by default. This framework fixes that.
It gives you a ready-to-use method to:
- Configure an AI that mimics your voice (tone, vocabulary, rhythm, values),
- Eliminate repetitive prompting (no more re-explaining everything at the start of each conversation),
- Maintain quality over time (preventing the AI from drifting toward generic language),
- Fix mistakes in a few minutes per month.
The result: a digital assistant that produces texts ready to send (or needing less than 2 minutes of touch-up), 100% in your style.
- Freelancers, consultants, experts whose "voice" is their brand,
- Business owners who want to delegate communication without losing their identity,
- Sales reps, marketing managers who write content daily,
- Anyone who uses AI and is tired of generic, robotic responses.
No technical skills required. All you need is an account on an AI platform that supports the "Project" feature (Claude Pro or ChatGPT Plus), and about thirty of your own texts.
💡 What is a "Project" feature? Some AI platforms let you create a dedicated workspace called a "Project" (or similar). Inside this workspace, you can set permanent instructions that the AI will follow for every conversation — without you having to repeat them each time. Think of it as a "briefing sheet" permanently attached to your AI's memory. Claude Pro calls this "Projects"; ChatGPT Plus calls it "Custom Instructions" or "Projects" depending on the version.
💡 What is a repository? A repository (often called a "repo") is a folder hosted on GitHub that contains all the files for a project. GitHub is a website where developers (and increasingly non-developers) store, share, and collaborate on files. Think of it as a Google Drive, but built for code and text files, with version history included.
| File / Folder | What it contains |
|---|---|
templates/Prompt_Master.md |
The Constitution: style rules, tone, forbidden phrases |
templates/Meta-Prompt_Nettoyage.md |
Extracts the 7 best example texts from your collection |
templates/Golden_Dataset.md |
5 regression tests — monthly check against voice drift |
templates/Quarantaine.md |
Daily log of AI mistakes (zero friction, copy-paste only) |
templates/Changelog.md |
Change log — tracks every configuration update with dates |
docs/ |
Complementary guides (glossary, quick start, architecture) |
examples/ |
Concrete examples of emails and posts |
💡 What are
.mdfiles? Files ending in.mdare Markdown files — plain text files with light formatting (headings, bold, bullet points). You can open them with any text editor (Notepad, VS Code, etc.) or read them directly on GitHub, where they render beautifully. No special software is required.
📂 custom-ai-assistant-framework/
├── 📄 README.md
├── 📄 LICENSE
├── 📄 .gitignore
├── 📁 docs/
│ ├── 📄 architecture.md
│ ├── 📄 glossaire.md
│ └── 📄 guide-de-demarrage.md
├── 📁 examples/
│ └── 📄 email-prospection-exemple.txt
└── 📁 templates/
├── 📄 Changelog.md
├── 📄 Golden_Dataset.md
├── 📄 Meta-Prompt_Nettoyage.md
├── 📄 Prompt_Master.md
└── 📄 Quarantaine.md
- Recommended: Claude Pro (the most advanced "Projects" interface).
- Alternative: ChatGPT Plus ("Projects" or "Custom Instructions" feature).
💡 Both platforms require a paid subscription (approximately $20/month). Free plans exist but do not include the persistent "Project" feature, which is essential for this framework to work.
- Open your email inbox, CRM, LinkedIn posts, or social media.
- Select 30 to 50 texts you wrote yourself (emails, posts, proposals, client replies).
- Don't over-filter: just pick what feels representative of how you naturally communicate.
- Paste them all into a single text file, one block per text, separated by blank lines or numbered.
💡 Why 30 texts? The AI needs enough examples to identify the patterns in your writing — your sentence length, your preferred transitions, your vocabulary, your level of formality. Ten texts is too few; a hundred is overkill. Thirty is the practical sweet spot that gives the AI enough material to work with without overwhelming it.
- Open
templates/Meta-Prompt_Nettoyage.md. - Copy its entire content.
- In the AI, start a new conversation, paste this prompt, then attach your file of 30–50 texts.
- The AI will select the 7 texts most representative of your voice, and for each one generate a Stylistic Identity Card (DNA + Anti-patterns).
- Save the result in a file called
Exemples.md. This is your Case Law.
💡 What is a "prompt"? A prompt is simply the instruction or question you give to an AI. It can be one sentence ("Write me a sales email") or a very detailed document (like the templates in this framework). The more precise and structured the prompt, the more predictable and useful the AI's response.
💡 What is a "Stylistic Identity Card"? When the AI analyzes your texts, it produces a short profile of each one: what makes it sound like you (your "DNA" — recurring phrases, rhythm, tone) and what to avoid ("Anti-patterns" — things that would signal the text was written by someone else or by a generic AI). Think of it as a writing fingerprint.
💡 What is "Case Law"? This metaphor borrows from law: just as judges refer to past rulings to ensure consistency, your AI will refer to your best example texts to ensure all future outputs stay consistent with your voice. Your
Exemples.mdfile acts as the precedents.💡 How do you "attach a file" to an AI? Most AI platforms with a "Project" or chat interface have a paperclip icon or an "Upload" button that lets you attach a file to your message. The AI will read its contents and use them as context for its response.
- Open
templates/Prompt_Master.md. - Replace all the text between
[brackets]with your own words:- Your profession, your target audience,
- Your 3 tone pillars (e.g., "Radical candor", "Pragmatic empathy", "Clarity"),
- Your forbidden phrases (e.g., no "I understand your frustration", no "In conclusion").
- Save this personalized file.
💡 Why call it a "Constitution"? Just like a country's constitution is the supreme document that all laws must respect, your
Prompt_Master.mdis the supreme document that governs everything your AI says. Every response it generates will be filtered through these rules. It's the single most important file in the whole framework.
- In the AI interface, create a Project (or equivalent).
- In the system instructions field, copy and paste the content of your personalized
Prompt_Master.md. - In the knowledge base, upload your
Exemples.mdfile. - You're done! Start asking questions. The AI will respond in your voice.
💡 What are "system instructions"? System instructions are a special field in a Project where you write instructions that the AI silently applies to every single conversation within that project — without you having to repeat them. It's the invisible briefing that runs in the background at all times. This is fundamentally different from typing instructions in a normal chat message.
💡 What is a "knowledge base"? In Claude's Projects interface, the knowledge base is a document storage area where you upload reference files. The AI can access these files during any conversation in the project, as if they were always open on its desk. This is where you put your
Exemples.mdso the AI always has your writing samples at hand.
To make sure your assistant stays faithful over time — AI platforms update silently, and behaviors can shift:
- Open
templates/Golden_Dataset.md. - In your Project, start a conversation named
🧪 REGRESSION TEST. - Run the 5 tests one by one, and check that the responses meet the success criteria.
- If a test fails, adjust your Master Prompt or your Examples (see
docs/guide-de-demarrage.md).
💡 What is a "regression test"? In software development, a regression test checks that something which used to work still works after a change. Here, the idea is the same: once a month, you run 5 specific writing scenarios through your AI to verify it still sounds like you. If it starts sounding generic or corporate again, you know the configuration needs a tune-up.
💡 What does "AI drift" mean? AI platforms are updated regularly by their developers. These updates can subtly change how the AI interprets instructions, which may cause it to gradually sound less like you and more generic. "Drift" is this slow, silent deviation from your intended voice. The monthly test is your early-warning system.
It will happen occasionally. Here's what to do:
- Correct the text manually before sending it (professional standards apply).
- Copy-paste the failed prompt and response into
Quarantaine.md(on your local computer). - At the end of the month, run a "Doctor session": give the AI your
Quarantaine.mdfile and ask it to identify the 3 changes to make to your configuration. - Log those changes in
Changelog.mdand update your files accordingly.
💡 What is a "Quarantine" file? It's a simple running log where you paste examples of the AI's mistakes — like a notepad for bugs. Zero friction: no formatting required, just copy-paste. At the end of the month, this raw collection of failures becomes the input for a structured improvement session. The AI itself helps you diagnose what went wrong and how to fix it.
💡 What is a "Changelog"? A changelog is a running record of every change made to a configuration or software, with dates and descriptions. It lets you roll back if a change made things worse, and helps you remember what you tried. Think of it as a diary for your AI setup.
docs/glossaire.md— Technical vocabulary explained in plain language.docs/guide-de-demarrage.md— An even more detailed step-by-step guide.docs/architecture.md— How the system works internally.
This repository is licensed under the MIT License (see LICENSE).
You are free to use it, modify it, and redistribute it.
Use it at your own risk: this system is an assistant, not a substitute for your professional judgment. Always read
texts before sending them.
💡 What is an MIT license? The MIT license is one of the most permissive open-source licenses. It essentially means: "Do whatever you want with this — use it, modify it, sell it — just keep the original copyright notice." It imposes almost no restrictions on the user.
This framework is the result of rigorous stress-testing, designed to be pragmatic, robust, and bullshit-free.
If you have questions, suggestions, or want to adapt this framework to a specific industry, feel free to open an issue on this repository or contact the author via GitHub (@valorisa).
💡 What is an "issue" on GitHub? An issue is a comment thread attached to a repository. It's the standard way to report a bug, ask a question, or suggest a feature. You need a (free) GitHub account to open one. Think of it as a public forum attached directly to the project.
Happy configuring! 🚀