|
| Context Is King This week's newsletter returns to some of the fundamentals. Many people I talk with about AI are using it every day but are still missing the basics of context management. They open a chat, ask a question, get an answer, and move on.
Yes, there's global memory that carries some context between conversations, but it thins out over time as unrelated topics dilute what AI remembers. Need to reference something from a month ago? It's probably been pushed out of global context by everything else you've discussed.
Well, there's a better way. Not only can you manage this global memory, but you can add your own custom instructions that apply to every chat you create. Great, right? Yes. But there's more!
The real power most people overlook are the "project-like" features: Custom GPTs, Claude Projects, Gemini Gems. These allow you to keep separate sets of context per project, and they remain isolated and don't get diluted by conversations in other projects.
If you're using AI and haven't set these up, you're working harder than you need to. These aren't advanced tools. They're basics that many people haven't explored, and using them will absolutely improve your results, guaranteed. ;) | Personal context settings in AI chatbots let you define who you are, what you do, and how you like responses. Set them once, and every future conversation starts with the AI already knowing these details. | Adding more prompt details often makes AI worse. Providing more instructions when something goes wrong creates prompts that can degrade performance. Clear and concise beats long and bloated. | Hidden system prompts shape every AI response before your chat begins. This explains why AI chatbots feel different even with similar models, and why your prompts sometimes seem overridden. |
|
| Introduce Yourself to Your ChatbotEvery major AI chatbot now lets you set personal context that persists across all your conversations. ChatGPT calls it "Custom Instructions," Claude offers "Profile Preferences," and Gemini has "Your Instructions." The concept is the same: tell the AI who you are, what you do, and how you prefer responses, and it will remember for every future chat. No more repeating that you're a marketing director, that you prefer concise answers, or that you work in healthcare. Setting this up takes five minutes and immediately improves every interaction. Include your role, industry, and common tasks. Add preferences for tone and format. Specify constraints like "avoid jargon" or "assume technical knowledge." These instructions layer on top of the chatbot's built-in system prompts, giving you consistent, relevant responses without the repetitive setup. Start simple. A few sentences about who you are and how you like information delivered will make a noticeable difference. You can refine over time as you notice what's working and what needs adjustment. ChatGPT's Custom Instructions → Claude's Profile Preferences → Gemini's Your Instructions → |
Adding more rules to fix AI failures often backfires. Prompts are effective for shaping tone, structure, and task framing, but they can't enforce accuracy or prevent hallucinations. The instinct to keep adding instructions when something goes wrong creates bloated prompts that actually degrade performance. The model spends more effort parsing your rules than doing the actual work. Practical Advice: Clear prompts beat long prompts. If your conversation is still going sideways after a few exchanges, start fresh with a clean prompt rather than layering on more corrections. Often the best fix is a reset. Read the Article → Every time you chat with an AI, invisible instructions are shaping how it responds. These system prompts define the AI's personality, safety boundaries, and formatting rules before your conversation begins. This is why ChatGPT, Claude, and Gemini can feel so different, even when running on models with similar capabilities. Strategic Takeaway: Your custom instructions are added on top of these baked-in system prompts. This hierarchy can explain why the same request produces different results across platforms, and why your personal instructions sometimes feel overridden. Read the Article → | Quick HitsBuilding Context Through Experience When an LLM's output isn't right, don't just fix it and move on. Notice what's wrong, get specific about why, and then ask the LLM to write up the patterns so you can use them in future conversations. Each correction becomes personal context that compounds over time. | One AI, Many Contexts (Custom GPTs, Projects and Gems) Wouldn't it be nice to create and store AI context for different tasks? You can. All three major platforms let you define instructions and upload files in separate spaces. That's pre-built context. Create one for your job, another for learning guitar, a third for working on your car. Each has its own context and won't get diluted by unrelated conversations. | Foundations Why Context Changes Everything in AI The sentence "I didn't say you stole my recipe" has seven different meanings depending on which word you emphasize. AI faces the same challenge: without context, it can't know which interpretation you need. General AI knows how to bake cookies, but context tells it you're at high altitude, your daughter has a dairy allergy, and you're baking for 30 kids. That's the difference between a generic recipe and a useful one. Better context equals better results. |
|
| Industry DevelopmentsMoltbook: A Social Network For AI Agents With over 1.5 million agents (and counting) posting across thousands of forums, Moltbook is a social network where bots are "exchanging information" at an incredibly rapid pace. It's an early glimpse of what AI agent collaboration at scale might look like, including very real PII and security risks. | Kaggle: A Gaming Arena to Benchmark AI's "Soft Skills" Adding Werewolf (a light RPG) and poker alongside chess, Google's Kaggle Game Arena tests AI models by measuring how they handle imperfect information, social deduction, and calculated risk. Gemini 3 Pro and Flash currently lead all three leaderboards. |
|
| |
|
