# Build Your Own AI Assistant Without Code: A Template

> To build your own AI assistant, set up a custom GPT or project in ChatGPT, a project in Claude, a Gem in Gemini or an agent in Copilot, then add standing instructions and reference files. The model isn't retrained; it reads your setup at the start of every chat. Precise instructions, clean files and a ten-question test decide whether it works.

- URL: https://dein-privater-assistent.de/en/tools/build-your-own-ai-assistant/
- Language: en
- Updated: 2026-10-19
- Publisher: Dein privater Assistent (OnlineBuilders LLC, https://onlinebuilders.biz)

If you type the same reminders into a chat every day ("write like us", "our prices are…", "keep it informal"), you're doing too much work. Every major assistant now lets you save that guidance permanently. The result is your own assistant that knows your services, your tone and your rules.

No coding required. One misunderstanding is worth clearing up first, though: the model isn't "trained on your data". It reads your instructions and files at the start of each conversation, like a colleague who glances at the briefing folder before every meeting.

## What a custom assistant is, and what it isn't

Technically it has two parts. Instructions that apply before every chat, and reference files the model consults when needed. Some tools add actions, meaning connections to other services, but most people don't need them.

What it isn't: a system that automatically reaches into your accounting software, CRM or inbox. That takes integrations beyond what these tools offer out of the box.

## Four ways to do it

**Custom assistants at the major vendors (October 2026)**

| | ChatGPT | Claude | Gemini | Copilot |
|---|---|---|---|---|
| Called | GPTs, Projects | Projects | Gems (skills from Nov.) | Agents |
| Saved instructions | yes | yes | yes | yes |
| Reference files | yes | yes | yes | partly |
| Usable for free | partly | yes, up to 5 | yes | partly |
| Public sharing | yes | no | partly | no |

"Partly" depends on plan or licence. Copilot Chat includes simple agents based on instructions and public websites.

Source: Help pages from Anthropic, Google and Microsoft, October 2026

ChatGPT has two forms: custom GPTs, which you can share, and projects, which work more like a personal workspace. GPTs are in transition, though. OpenAI's [developer documentation](https://learn.chatgpt.com/codex/migrate-custom-gpts) says custom GPTs are being replaced by plugins, without giving a retirement date (as of October 2026). Keep a text copy of your instructions and files here too. Claude centres on projects. According to [Anthropic's help centre](https://support.claude.com/en/articles/9517075-what-are-projects), free users can create up to five, and paid plans add retrieval that stretches file capacity by up to ten times.

With Gemini, check the calendar. [Google's help page](https://support.google.com/gemini/answer/15146780?hl=en) says that from November 2026, Gems on personal Google accounts go away and transition automatically to skills. If you build a Gem now, keep a copy of the instructions as a text file.

Microsoft's Copilot documentation lists simple agents grounded in instructions and public websites as part of Copilot Chat. Agents that use company data are billed by usage or need the premium licence.

## The instruction template

Instructions are the heart of it. Here's a template for a quoting assistant at a small tiling business, ready to adapt:

```text
Role: You help the office at Harlow Tiling Ltd draft quotes
and customer emails.

Audience: homeowners, not trade professionals.

Knowledge: Use only the files "Services.pdf", "Prices_2026.pdf"
and "Style.txt". If a price isn't listed there, write [PRICE MISSING]
instead of estimating.

Tone: friendly and plain. Short sentences. Explain any trade term.
No superlatives.

Process: If room size, preferred material or timing are missing,
ask for them first. Only then write the draft.

Limits: No statements about warranties, structural questions or
grants. Refer those to Tom Harlow.
```

Three lines carry most of the weight. "Use only the files" stops the model pulling prices from its training data or the web. "[PRICE MISSING] instead of estimating" makes gaps visible. And the limits keep it away from statements you'd be liable for. Our guide on [how to write AI prompts](https://dein-privater-assistent.de/en/tools/how-to-write-ai-prompts/) covers more phrasing like this.

## Three assistants that almost always pay off

The template shows a quoting assistant. Two other types work well in small companies because they combine repetitive work with documents that already exist.

The content assistant gets a style sample of ten published posts, a list of words you never use, and a description of your audience. Its job is turning bullet points into a newsletter or social post draft. Someone still approves it. Anyone publishing AI-assisted content should also know when a label is required; our piece on [disclosing AI-generated content](https://dein-privater-assistent.de/en/privacy-and-limits/disclose-ai-generated-content/) covers that.

The onboarding assistant knows the staff handbook: processes, contacts, where each template lives. New colleagues ask it before interrupting the same person for the third time. The key rule here: "If the answer isn't in the handbook, say so and name the person responsible." Otherwise it invents processes that don't exist.

All three share two traits. They work from material you already have, and someone checks their output before it leaves the building.

## Which files belong in the knowledge base

Good reference files are short, current and unambiguous. Price lists as tables beat a 40-page catalogue full of photos. An FAQ built from real customer questions beats a brochure.

What doesn't belong: client data, contracts with names, staff lists, passwords. Doubly so if you share the assistant. With publicly shared GPTs, it's well known that users can extract instructions and file contents with clever questions. For the legal side of personal data, see our [ChatGPT and GDPR guide](https://dein-privater-assistent.de/en/privacy-and-limits/chatgpt-gdpr/).

One practical tip: add a style sample as its own file. Three or four real emails you like, headed "This is how we sound." That does more than any description of your tone. Our [AI email assistant guide](https://dein-privater-assistent.de/en/everyday-automation/ai-email-assistant/) shows what that looks like for email.

## Instructions or memory?

ChatGPT and Claude also have a memory feature that remembers individual facts from earlier chats. Handy, but hard to control: you only see what was stored if you go and look. Anything that must apply reliably belongs in the saved instructions. Memory is for small things like your preferred formatting, not prices or rules.

## The ten-question test

An untested assistant is a liability. Spend half an hour asking it ten questions that could genuinely come up.

**Ten test questions every assistant should pass**

- [ ] Three typical requests it should handle well
- [ ] One request with missing information (does it ask?)
- [ ] A product or price not in the files (does it invent one?)
- [ ] A question outside its job (does it decline?)
- [ ] A warranty or legal question (does it refer on?)
- [ ] A rude customer (does the tone hold?)
- [ ] A very long request with several asks
- [ ] A follow-up to an earlier answer in the same chat

For each mistake, add a rule or fix a file, then rerun the test. Keep the ten questions in a text file. When you change the instructions later or the vendor rolls out a new model, ask the same questions again and you'll see straight away whether anything got worse. After two or three rounds most assistants are fit for daily use. Why models tend to invent rather than ask when information is missing is explained in our piece on [AI hallucinations](https://dein-privater-assistent.de/en/privacy-and-limits/ai-hallucinations/).

## Maintenance, the part everyone forgets

Prices change, services too. An assistant holding the spring price list writes wrong quotes in autumn, and does so convincingly. Put a fixed date in your calendar, say the first of each month, to review the files. If your team runs several assistants, make one person responsible.

Putting a date in file names helps, for example "Prices_2026-10.pdf". You can see at a glance how old the information is, and the assistant can tell users which version it's quoting from.

## When built-in tools stop being enough

Custom GPTs and projects hit their limit once the assistant needs live data: stock levels, open invoices, free appointment slots. Then it has to connect to your systems, with proper permissions and logging. That's development work, the kind of chatbot and virtual assistant build [OnlineBuilders](https://onlinebuilders.biz/solutions/ai-bots) does for companies.

Until then, the next step is modest. Write down the three reminders you type into chats most often and make them the first lines of your instructions. Details on individual tools are under [tools](https://dein-privater-assistent.de/en/tools/).

## Step by step

1. **Define the job** Write one sentence on what the assistant is for and what it isn't, for example drafting quotes for bathroom refits but no legal questions.
2. **Pick the tool** Use what you already have: a custom GPT or project in ChatGPT, a project in Claude, a Gem in Gemini, an agent in Copilot.
3. **Write the instructions** Set out role, audience, tone, process and limits. Start from the template below.
4. **Upload reference files** Price list, FAQ, style sample, service descriptions. No client data, nothing confidential.
5. **Test with ten real questions** Ask everyday questions, including two the assistant should decline. Note every mistake.
6. **Tighten and retest** Add a rule or fix a file for each mistake, then run the test again.

## FAQ

### Does my assistant learn from conversations?

Not really. It rereads your instructions and files in every chat. Some tools also have a memory feature that stores individual facts, but improvements mostly come from you tightening the instructions.

### Can I share my assistant with clients?

Custom GPTs can be shared by link or in the GPT directory. Bear in mind that determined users can often coax out the instructions and file contents. Don't put anything in there that mustn't become public.

### What's happening to Gemini Gems?

Google's help centre says that from November 2026, Gems on personal Google accounts will transition to something called skills, automatically. Save your instructions as a text file beforehand anyway.

### How many files can I upload?

That depends on the vendor and plan. For Claude Projects, Anthropic lists 30 MB per file, with the total needing to fit in the model's context; paid plans add retrieval that stretches capacity.

Photo: Oli Dale olidale, CC0 (https://commons.wikimedia.org/wiki/File:Ruler_and_laptop_on_a_desk_(Unsplash).jpg)