Why a contextless chat always gives you weak answers
A language model answers with the common knowledge in its training data. Ask for social media copy without giving it framing, and it returns something generic, because it doesn't know who you are or what you sell. Most people blame the machine, but the problem is almost always missing context in the prompt.
Most people blame the machine when the answer is weak, but the problem is almost always the lack of context in the prompt. A good question brings a good answer; a bad question brings a bad answer. The machine cannot read your mind. You have to write it down.
When you work with a custom assistant instead of a generic GPT opened in a new tab, the baseline instructions are already there before you ask anything. It knows who you are, what kind of customer you serve, which words to avoid. That changes the daily routine: instead of writing a huge prompt every new conversation, you go straight to the point and the answer is already aligned with your business.
There is also a practical difference between creating a custom GPT and creating a project inside ChatGPT, which are essentially similar in function. The most relevant difference is that a custom GPT can be shared with other people or teams, while a project tends to stay more confined to whoever created it. I keep both set up in parallel with the same information loaded, but in daily use I end up relying more on the custom GPT.
The behavioral instructions that tell the AI how to speak
General behavioral instructions set the baseline for how the assistant always responds before any specific question. Here you define the language variant and the tone for your interactions. You also explicitly ban writing tics that instantly give away machine-generated text to ensure cleaner, more natural responses throughout.
Since I work for the Portuguese market, I explicitly request European Portuguese and ban em dashes and other tics that show up constantly in AI-generated text. If your tone is direct, without detours, write that into the instructions: something like "I don't like flowery language in responses" is a simple line that already steers the output a lot.
You can also assign the assistant a functional role, such as an experienced digital marketing professional focused on results, so it reasons from that frame instead of giving unmoored answers. This helps particularly when you ask for analysis or opinions, not just ready-to-publish text.
An instruction I use and that is worth considering: ask the assistant that, whenever it identifies something relevant in a conversation, something you told it or a conclusion drawn from a real result, it suggests at the end of its answer a short text you can paste into its own instructions. That way it keeps updating over time based on what actually happened in your work, instead of staying frozen at the version you first uploaded.
How to Give AI Context Without Dumping Cluttered Information
How to use AI in business beyond chatbot prompts
How to plan an app with Claude Code before you start coding
Brand content: personas, real fears and objections
What the assistant knows about your audience should not come from the model's imagination, it has to come from whoever deals with the customer daily, with the detail only that experience gives. If you give it a shallow persona, it will generate ideas based on stereotypes that don't convert.
It is worth documenting your customer's fears, the most frequent objections your sales team hears, their desires, and what differentiates you from competitors. It also helps to list unshakeable beliefs about the business, things like ideas without action are just dreams, or focusing your energy on what you control, because that gives the assistant a compass when it has to choose between several possible directions in an answer.
This isn't work to delegate blindly to AI: start by writing the examples you already know from your own business yourself, then use the machine to help expand the list, always checking whether it fits your real audience and isn't just a generic marketing exercise that could apply to any company.
Metrics and business limits the machine should respect
If the assistant doesn't know your priority metrics, it risks suggesting ideas that don't make financial sense for you. It's worth writing down that your main KPI is profit, or customer acquisition cost, or lifetime value, so any suggestion it makes takes that into account.
Documenting the traffic sources you've already tried, what worked and what didn't in past campaigns, also helps the assistant avoid repeating suggestions you already know, from practical experience, don't work for your specific case. If you know a certain social network delivered weak results in the past, write that down, so the machine doesn't keep suggesting you invest there with no track record backing it.
This kind of information is also useful when you ask the assistant to comment on a briefing or a strategy. The more it knows about real priorities, available team, and tools already in use, the closer the answer gets to something applicable, instead of a theoretical idea that sounds nice but is hard to execute with the resources you have. Writing these instructions well is not a five-minute job, and it's not worth asking the AI itself to write them for you, because this information is too specific to your own business.
Organizing files without drowning the assistant in information
In the class I showed how to add reference files to an assistant. Select those documents carefully: old terms, conflicting versions and examples taken out of context can undermine the response. Choose the material needed for the task and review it when operations change, respecting the limits of the tool you are using.
If you have a lot of content, like transcripts from several work sessions, it's worth compressing several files into one organized document by thematic blocks, instead of uploading everything scattered at random. And most important: only upload what is actually relevant to the assistant you're building. If it's just filler, don't upload it, because that only introduces noise and can make the machine favor the wrong source.
You can also turn off web search in the assistant's settings, forcing it to answer only based on what you gave it. This avoids it mixing your information with generic answers pulled from the web, and is especially useful when you want it to function as an extension of your own way of thinking, not a disguised search engine.
A different assistant for each client or project
When working with multiple clients, as happens in an agency, it doesn't make sense to use the same assistant for all of them. Each client has different positioning, audience, and offer, and mixing it all into the same context space just confuses the answers, risking one client's tone bleeding into another client's copy.
The sensible approach is to create a dedicated assistant for each client or project, loaded with that account's onboarding information, tone of voice, and specific digital assets. That way, when someone on the team needs to work on that client, they use the right assistant, without the risk of mixing language or goals across different accounts.
This separation also makes maintenance easier over time. Each assistant gets updated as you learn more about that specific client, instead of having one generic assistant where it's impossible to tell which information belongs to which account.
Source
This article is based on the live session Sextas Ímpares #119: Como CRIAR E TREINAR um Assistente de AI no ChatGPT, which I recorded and shared on YouTube. In that live I showed in real time how to configure the instructions and file base of a custom GPT, explained the practical difference between creating a GPT or a project, and answered questions from people following along live. You can watch the full session here: https://www.youtube.com/watch?v=bIsuNjYwi8k. It's worth watching from [00:19:56] for the explanation on how to write good prompts, from [00:38:16] for the moment I show, in practice, where instructions get pasted inside the GPT configuration, and from [01:26:41] for the part about creating simple automated tasks inside the assistant itself.