A custom GPT is a version of ChatGPT that your company configures for its own needs. To create a custom GPT, you give it two things:
- instructions, also called the system prompt: a text you write, which the GPT applies in every conversation;
- a knowledge base: the documents you upload to the GPT, which it consults when a question concerns their content.
The instructions describe the GPT’s role, your company, your readers and your writing rules. These rules come from two decisions: what you sell, and how your brand talks about it. The knowledge base holds the facts the GPT must refer to: the description of your offers, your results, your proof. If your writing rules sit in the knowledge base instead of the instructions, or if one of the two is missing, the GPT produces copy that any competitor could sign.
ChatGPT’s competitors offer the same function under other names: Projects in Claude, Gems in Gemini, agents in Microsoft 365 Copilot. The advice below applies to these three as well.
Your teams already write with AI, but with tools your company has not configured
In May 2024, Microsoft and LinkedIn published their annual survey on work, the 2024 Work Trend Index. The research firm Edelman surveyed 31,000 knowledge workers, full-time employees or freelancers, in 31 countries, between February and March 2024. According to the survey, 78% of people who use AI at work bring their own tools. They do so, the survey adds, without guidance or approval from their management.
For a marketing or communications department, this 78% has a direct consequence. Most of the copywriters, product managers and sales staff who write in your brand’s name with AI use a tool your company has not configured. That tool knows neither the description of your offers, nor the words your brand uses, nor the words it rejects. It produces correct but generic copy.
A custom GPT corrects this on one condition: your company has put its own writing rules and the facts about itself into it. The example below compares two assistants that did not receive the same documents.
An example: two assistants write for the same accounting firm
On 29 May 2026, WeAreTheWords ran the AI Editorial Assistant workshop, a one-day session in French on configuring a writing assistant. For the demonstration, we invented a brand: Au Carré, a French accounting firm for the self-employed. We created two assistants in Claude, Anthropic’s AI, and asked both to write on the same subject: moving from a micro-entreprise, France’s simplified regime for sole traders, to a company. The table sets out, side by side, what each assistant received, what it wrote and what its text contains.
| First assistant | Second assistant | |
|---|---|---|
| What the assistant received | One document: the firm’s editorial charter, pasted as it stood in place of instructions. | Five items: the description of the firm’s offer, the firm’s writing rules, a list of banned words, a glossary of accounting terms and a list of checks to run before returning a text. |
| What the assistant wrote | “At a time when many self-employed people are asking themselves the question, becoming a company is a key step in your development. At Au Carré, your trusted partner, we support you with peace of mind in making the choice best suited to your needs and ambitions.” | “Moving from a micro-entreprise to an EURL or a SASU is almost never a question of image. It’s a question of costs. As long as your actual costs stay below the flat-rate allowance, the micro-entreprise suits you. Above that, we look at the figures together.” |
| What the assistant’s text contains | No information specific to the firm. This text could appear on any accountant’s website. | The two possible legal structures, the EURL (single-member limited liability company) and the SASU (single-shareholder simplified joint-stock company); the decision criterion (actual costs compared with the micro-entreprise’s flat-rate allowance); what the firm offers when costs exceed the allowance: reviewing the figures with the client. |
What the example shows. An editorial charter alone is not enough to configure an assistant. Without the description of the offer or the list of banned words, the first assistant filled the gap with general phrases such as “trusted partner”.
Yet the participants in the 29 May workshop did not have the documents given to the second assistant. There were eight of them, from four sectors: retail, business services, digital agencies and social care. On arrival, none had a written description of their offer. None had writing rules drafted so that an assistant could apply them either. To configure their assistant, each participant had to make it infer their brand’s tone from texts already published. Configuring a custom GPT therefore requires work beforehand: deciding what you sell, then describing how your brand talks about it.
The instructions: how to write the system prompt
The instructions apply to every conversation opened in the GPT, whatever text is requested. They cover more than writing rules. To write them, you can start from the prompting best practices published by Anthropic, the company behind Claude. Among other things, this guide recommends:
- being clear and direct;
- giving the AI a role;
- adding context;
- providing examples;
- specifying the expected format.
From these recommendations, we defined eight sections for the instructions of a writing assistant. Each section is illustrated with an example written for this article, using the fictitious firm Au Carré.
| Section | What you write in it | Example for the fictitious firm Au Carré |
|---|---|---|
| Role | The GPT’s function in the team. | “You are the copywriter for Au Carré.” |
| Context | Your company, its offers and its clients, in a few sentences. | “Au Carré helps self-employed people choose their legal structure and keep their accounts.” |
| Tasks | The texts the GPT writes, and what it does not do. | “You write the blog articles and the emails to clients. You never give tax advice with figures.” |
| Readers | The intended readers, and what they already know about the subject. | “Your readers are micro-entrepreneurs who do not know accounting vocabulary.” |
| Approved terms and banned words | The name of each offer, the words that change the meaning of the offer, the stock phrases your brand rejects. At the 29 May workshop, we recommended a list of 30 to 60 banned words. | “Never use: ‘trusted partner’, ‘support you with peace of mind’.” |
| Tone | How to address the reader, and the level of technical detail. | “You address the reader in a courteous, professional register. You explain each technical term the first time you use it.” |
| Model text | A text already approved, which the GPT takes as its model. | A blog article approved by the firm, pasted in full. |
| What to do when in doubt | What the GPT does when a piece of information is missing. | “If a figure is not in your documents, say so instead of inventing it.” |
The approved terms, the banned words and the tone come from two decisions taken before configuring the GPT:
- the offer: what you sell, to whom, and the promise you keep. Our article on offer architecture explains how to state that offer;
- the voice: your brand’s words, tone and narratives.
At WeAreTheWords, when a client company has not yet decided the promise of its offer, we do not configure its assistant. That assistant’s instructions would contain no approved term and no promise to use. It would produce general phrases, like Au Carré’s first assistant.
The knowledge base: what goes in and what stays out
The GPT does not read its whole knowledge base for every answer: it searches it for the passages relevant to the question asked. The knowledge base therefore suits facts, which serve only for some questions. Writing rules apply to every answer, so they go in the instructions. The following table shows, for each document, whether it belongs in the knowledge base, and why.
| Document | In the knowledge base? | Why |
|---|---|---|
| The detailed description of each offer: content, target audience, how it runs | Yes | Whenever a text presents an offer, the GPT consults this description. |
| The results you achieve for your clients, with figures where possible | Yes | These results provide the material for any text that must show what an offer delivers. |
| Proof: client cases, testimonials published with their author’s consent, figures with their source stated | Yes | The GPT can cite only the proof present in the knowledge base. |
| Writing rules | No | Stored in the knowledge base, a writing rule applies only to the questions that lead the GPT to consult the passage containing it. Writing rules therefore go in the instructions. |
| Documents that contradict each other | No, only the current version | If the knowledge base holds two versions of a brochure with two different promises, the GPT repeats the first promise in some answers and the second in others. |
| Texts already written by AI | No | The GPT would carry the general phrasing of these texts into your new publications. The article “Why AI content is bland: the cause is mathematical” explains this mechanism. |
| Confidential documents | No | Anyone you give access to the GPT can ask it to display the content of the files uploaded to the knowledge base. |
Once you have chosen the documents, record their origin. For each one, note its date, the department that produced it and the person who approved it. When the GPT makes a claim, you can then find the document it comes from, and correct that document rather than the answer. When a document changes, replace the old version instead of adding the new one alongside it.
Testing your custom GPT before handing it to your teams
Before sharing your custom GPT, give it the following four requests. Then put the same requests to standard ChatGPT, with no configuration, and compare the answers.
| Request | What your custom GPT’s answer must contain |
|---|---|
| “Summarise our offer in three sentences.” | The official sentence that presents the offer, with no banned word. |
| “Describe a typical situation for one of our clients.” | A client who resembles yours: their trade, their size, their problem. |
| “Back up our promise with proof.” | Only proof present in the knowledge base. |
| A question the knowledge base cannot answer, for example about a figure you have not put in it. | A statement that the information is missing, with no invented figure. |
A test passes when your custom GPT’s answer could not have come from standard ChatGPT. If the two answers look alike, correct your GPT’s instructions or its knowledge base: ChatGPT is not the cause.
After deployment, keep reviewing the texts the GPT produces. Our article on AI content supervision fatigue describes what happens when this review becomes mechanical. To set rules for AI use across the whole company, beyond this GPT, see our article on the editorial charter and AI.
Frequently asked questions about custom GPTs
Custom GPT, Claude project or Copilot agent: which should you choose?
Choose the tool your company authorises and your teams already use. The custom GPT, the Claude project and the Copilot agent work the same way: instructions applied to every conversation, and documents consulted when needed. Whichever tool you choose, the quality of the copy depends on what you write in the instructions and what you upload to the knowledge base.
What should go in a custom GPT’s knowledge base?
A custom GPT’s knowledge base contains the description of each offer, the results achieved for your clients and proof whose source is cited. Writing rules go in the GPT’s instructions. Texts written by AI and confidential documents stay out of the knowledge base. For each document, upload only the current version.
Who should keep a custom GPT up to date?
Assign the GPT to one member of the team. This person approves each document before it goes into the knowledge base and replaces outdated versions. When your offer changes, they update the GPT’s instructions first, then the documents that describe the offer.
WeAreTheWords’ AI editorial assistant applies these principles across the company: it writes only from your offer, your brand voice and your writing rules. If your offer or your voice is not yet written down, start with the Clarity engagement, which covers both. Let’s talk about your assistant.