If you use ChatGPT, Claude, or Gemini for work, you have probably noticed something annoying.
Every conversation starts from scratch, it forgets, goes off piste and often feels like it’s easier to do it yourself. So you spend the first five minutes of every chat explaining the same context.
It’s exhausting. If you miss something, you spend ages trying to get it right. What you may not realise is there is a way to “hard code” your LLM so that it can do specific jobs on repeat without the need for prompting.
CustomGPTs, Projects, and Gems fix this problem. They let you bottle up your context and reuse it. You set things up once, then every conversation starts with the AI already knowing what you need.
So let’s dig into the word of CustomGPTs, Gems and Projects…
What are they?
All three tools do the same job in slightly different ways.
They give you a way to save context and instructions so the AI behaves consistently across conversations.
- CustomGPTs sit inside ChatGPT. You build a custom version of ChatGPT for a specific task, give it instructions, upload reference documents, and then call it up whenever you need it. You can also share CustomGPTs with other people. There is a public store with thousands available.
- Projects in Claude are folders for your work. You create a Project, add instructions and reference documents, and every chat inside that Project inherits the setup. Conversations stay grouped together. Claude remembers context across chats inside the same Project.
- Gems are Google’s version, built into Gemini. You create a Gem with custom instructions and reference material. You can use it across Google Workspace tools.
The shared idea is simple. You package up your context once.
Why should you use them?
Three reasons.
- You stop repeating yourself. No more pasting the same brief, the same brand guidelines, the same audience profile, the same writing rules at the start of every chat. The setup is already there.
- You get more consistent output. When the AI has the same instructions every time, it produces work that sounds the same every time. That matters if you have a brand voice, a specific client, or a team using AI together.
- You build a system, not a habit. Most people use AI like a vending machine. Question in, answer out, repeat. CustomGPTs, Projects, and Gems turn AI into proper infrastructure. You build it once, it works for years.
To real value from AI you need to set it up properly. Junk in, junk out. Most type the same prompts every time, but often it will reproduce very different results. These ‘hard coded’ features help you get consistent output, but with so many applications you can use AI for, it’s important to know where feature like this sits.
So when is it worth building one?
When it is worth building one?
Not every task needs a custom assistant. A one-off question, a quick fact check or one off thing doesn’t need it. If you only do something once, a normal chat is fine.
Here’s a quick check for you:
- You do the same task every week: Drafting client emails. Writing social posts. Editing newsletters. Summarising meetings. Producing proposals. Anything repeated is worth setting up once.
- You have a specific voice or style to protect: If your writing has rules, your brand has guidelines, or your output needs to sound like you, a custom assistant keeps things consistent across every piece of work.
- You work with the same client, audience, or product over and over: A built-in assistant for a specific client means you stop briefing the AI from scratch every time. The context lives inside the tool.
- You have a team using AI: A shared assistant gives everyone the same starting point. The intern, the new hire, and the senior all produce work from the same brief.
- You keep getting generic output: If the AI keeps giving you bland, average, middle-of-the-road work, the problem is usually missing context. A custom assistant is the fix.
If none of those apply, use a normal chat.
Best practice for setting them up
Most people set these up badly. They write a vague paragraph of instructions and upload a random document. The output is generic and then they blame the tool. It’s important you give it guidelines and tracks to run on.
Here is how to do it properly:
1. Pick one job per assistant
The biggest mistake is building one CustomGPT to do everything. Marketing, sales, admin, ideas, research, all in one. The output ends up muddled.
Build one assistant per job. You can create them for almost anything, here is a list of ideas for you…
- Proposal writing
- Content drafting
- Prospect researching
- Feedback from sales calls
- Drafting emails and follow ups.
- Outlines for presentations.
If you can think of it, it has a structured setup, it can probably do it.
2. Write instructions like you are training a new hire
Imagine someone joined your team yesterday. They are smart but they know nothing about your business. What would you tell them?
Cover these things:
- Who you are and what your business does
- Who the audience or client is
- What the assistant should produce
- How it should sound
- What it should never do
Be specific. “Write in a professional tone” tells the AI nothing. “Write in plain English. Explain things like sentence length, readability, British English etc. It gives it expectations of how you want it to work.
Good instructions tell the AI what to avoid as much as what to do. If you hate certain phrases, rhetorical questions, you do not want bullet points unless you ask for them, say so in your instructions. You can even create a list of banned words, phrases and upload it all as knowledge.
3. Upload your best examples
Reference documents make a huge difference.
The AI learns your voice from examples better than from descriptions. Upload documents that give supporting context – this can be copies of emails, documents, technical inforation. This forms the context it draws upon for the output. Do not upload everything you have ever written. Upload your best examples of what good looks like.
4. Test it with real work
Once it is set up, run five real tasks through it. Tasks you would actually do this week. Then look at the output. If something is off, go back to the instructions and fix it. Keep tightening until the output is something you would send.
This part takes a few rounds. Treat it like training, not a one-off setup.
5. Update it when things change
Things change, but here is the good news. If your business changes, your audience changes or your offer changes – you can quickly update the knowledge and instructions too. You can either re-upload the knowledge and replace outdated information, or add more context.
It’s worth putting a calendar reminder in every quarter to review your CustomGPTs, Projects, and Gems. Update the instructions, refresh the examples and delete the ones you no longer use.
A simple starter setup
If you have never built one of these, start with something simple.
Build one first that will save you time. Get familiar with that one. Once you’ve mastered it once, it gets easier to do more and more. Pick one task you do every week whether that is writing, editing client emails, drafting social posts or summarising calls.
Build one assistant for that task. Use it for two weeks and improve it as you go.
Then build the next one.