Custom GPTs and Claude Projects: Building a Team Assistant
Practice4 minInnovialabs editorial team
In short
Custom GPTs and Claude Projects let you build an assistant from instructions and knowledge files in an hour; that is often enough for internal tasks, while integrations, access control and customer-facing work need a custom bot.
What it is and who it suits
Custom GPTs (in ChatGPT) and Claude Projects (in Claude) let you configure an assistant without code. You write instructions (how to answer, for whom, in what tone), upload knowledge files (policies, price lists, templates, FAQs) and share the assistant with colleagues.
It is a good option when the assistant is for employees, not customers: answering questions about internal rules, drafting emails and proposals from templates, checking texts against the company style, helping new hires in their first weeks.
Both features are available on certain plans, and team sharing depends on the plan. Check the current terms on the service's site before you start.
How to build the assistant, step by step
- Pick one task. "Answer sales team questions about delivery terms" works better than "help with everything".
- Write instructions. Role, audience, answer format, what not to do. For example: "Answer only from the uploaded files. If the answer is not there, say so and suggest contacting the head of the department."
- Prepare knowledge files. Clean text with clear headings works better than scans and slide decks. Remove outdated document versions, or the assistant will quote them.
- Add examples. Two or three samples of a good answer in the instructions noticeably even out the style.
- Test on real questions. Collect 15–20 typical questions from the team and check the answers, especially those where the right answer is "I don't know".
- Share and gather feedback. Assign an owner who keeps files and instructions up to date.
Example instructions
You are the sales team assistant. Reply in Russian or Uzbek,
in the language of the question.
Use only information from the "Delivery terms" and "Price list" files.
If the information is not there, say so. Do not make things up.
Format: a short answer, then a reference to the document section.
Never mention discounts that are not in the price list.
What works well and what does not
| Task | No-code assistant | Custom bot |
|---|---|---|
| Questions about internal documents | good | good |
| Drafting texts from templates | good | good |
| Talking to customers in Telegram or on a website | no, only inside the service | yes |
| Live CRM data | limited, via Actions or connectors | yes |
| Role- and department-based access | basic | flexible |
| Logs, analytics, quality control | limited | full |
| Time to launch | hours | weeks |
The main limits of the no-code approach:
- the assistant lives inside the ChatGPT or Claude interface and cannot be embedded in your product;
- knowledge file search works as is, with little room to tune it for your data;
- integrations are possible (for example, Actions in Custom GPTs call external APIs) but need technical setup and care with keys;
- every employee needs an account in the service.
Data access: what to check
- Plan and model training. Check your plan's settings and terms to see whether your data is used for training. Business plans usually have stricter terms, but verify your specific plan.
- Who can see the assistant. Share it within the team workspace, not via a public link. With a public Custom GPT, users may be able to extract the contents of its knowledge files.
- What to upload. Do not put customer personal data, passwords, financial statements or contracts with confidential terms into knowledge files without clearing it with whoever owns security.
- Personal accounts. If employees build assistants in personal accounts, the company loses control of the data. A central workspace is better.
When to move to a custom bot
Custom development makes sense when the assistant must talk to customers, work in Telegram or on a website, pull data from the CRM and accounting system, restrict access by role or keep full logs. The no-code assistant still makes a good prototype: it is a convenient way to test instructions and collect real questions.