The evolving landscape of personal AI assistants highlights the importance of tailored platforms and precise instructions

As the market for custom AI assistants becomes more complex, users must now navigate diverse platform capabilities and focus on defining narrow operating models for effective results.

Anyone trying to build a personal AI assistant now faces a more complicated market than many quick guides suggest. The big platforms all support some form of persistent instructions, saved context or reusable files, but they do not offer the same thing. Google now lets people create Gemini Gems on personal accounts, OpenAI separates reusable GPTs from its broader Projects workspace, and Anthropic has widened Claude Projects to free users as well as paid customers. For users choosing a tool, the important question is no longer whether custom assistants exist, but what sort of assistant each product actually provides.

Google introduced Gems on 28 August 2024 as part of a wider Gemini update that also brought Imagen 3, after previewing the feature at Google I/O. At launch, Dave Citron, Google’s senior director of product management for Gemini Experiences, described Gems as “custom versions of Gemini” and said users could “customize Gems to act as an expert on topics or refine them toward your specific goals”. That release was initially staged for Gemini Advanced, Business and Enterprise users over the following days. Google’s current Gemini help pages, however, show that the feature has since broadened: people using a personal Google account can now create and use Gems as long as they meet the age requirement and are signed in.

That matters because Gemini’s implementation is relatively straightforward. Google says users create a Gem by naming it, writing instructions and, if needed, adding files under a “Knowledge” section. The company also offers a built-in rewrite aid so Gemini can expand a rough description into fuller instructions. Once saved in the web app, the same Gem appears in the Gemini mobile app and the Gemini side panel in Google Workspace. Google also ships several premade versions, and TechRadar reported at launch that these included Learning Coach, Brainstormer, Career Guide, Coding Partner and Writing Editor. In practice, that makes Gemini the simplest route for someone who wants a repeatable expert-style chatbot without setting up a broader project space. There are limits, though: Google says custom Gems must be created, edited or deleted in the web app, and they cannot currently be used with Gemini Live.

OpenAI’s documentation draws a sharper distinction between different kinds of customisation. Its help centre says GPTs are purpose-built versions of ChatGPT that can combine instructions, conversation starters, uploaded knowledge and capabilities such as web search or image generation. That sounds closest to the classic idea of a bespoke assistant. But the same help page makes clear that “New GPT creation and publishing are not available on personal ChatGPT accounts, including Free, Go, Plus, and Pro.” In other words, a personal subscriber can use GPTs, but cannot currently create a new one unless they are working inside an eligible managed workspace.

For many users, that means ChatGPT Projects are the more realistic option. OpenAI now describes Projects as spaces where chats, files and instructions stay together for ongoing work, and says they are available to all free and paid subscription types globally. A project can keep its own memory, meaning later conversations inherit the same context instead of starting from scratch. OpenAI also says a chat moved into a project takes on that project’s instructions and file context. The product is therefore less like a public-facing custom bot and more like a structured work hub. It also supports tools already familiar from the main product, including Canvas, image generation, voice and web search. OpenAI has further expanded sharing, saying in an October 2025 update that project sharing became available across all ChatGPT account types.

Claude’s version sits somewhere between the two. Anthropic’s support material says Projects are available to all users, including free accounts, although free users can create only five. Inside a project, users can upload documents, text files or code snippets into a knowledge base that Claude will use across chats in that project. Anthropic also makes an important qualification that many casual guides miss: context is not automatically shared across separate chats unless the relevant material has been added to the project knowledge base. Users can also set project-wide instructions by selecting “Set project instructions”, after which Claude applies those instructions to all chats inside the project. On paid plans, Anthropic says Claude can automatically enable retrieval-augmented generation, or RAG, as a knowledge base approaches the context-window limit, expanding how much project material can still be searched. Team and Enterprise customers can also share projects and assign permissions.

Perplexity is harder to summarise in consumer terms because its own clearest guidance sits in developer documentation, but it still illustrates an important principle. The company’s Agent API prompt guide shows that a custom assistant depends on structured instructions, not only on improvised one-off prompts. One example tells the assistant to be “a concise, well-researched assistant” and to say so explicitly rather than guessing if searches fail. Perplexity also warns users not to rely on prose instructions to control search filtering, advising them to use built-in parameters instead for domains, dates and other search constraints. That is a useful reminder across all platforms: if a task depends on source quality, freshness or scope, the instruction set must be precise, and the platform’s native controls matter as much as the wording.

The practical lesson is that building a good assistant is less about writing a grand prompt and more about defining a narrow operating model. Start with the role, the task, the context and the output format. Then add boundaries: what sources are acceptable, what style to use, when the model should ask follow-up questions, and when it should admit uncertainty. Google’s own Gems guidance recommends focusing on persona, task, context and format, while Perplexity’s documentation underlines the value of explicit limits and clear failure conditions. A finance helper, for example, should not merely “help with budgets”. It should state whether it may estimate, whether it must cite uploaded spreadsheets first, and how to flag missing data.

That is why the best platform depends on the job. Gemini is now well suited to people who want a lightweight reusable assistant tied to a personal Google account. Claude Projects are stronger for document-heavy work that needs a durable knowledge base, especially where several chats should draw from the same files. ChatGPT Projects are best understood as persistent workspaces, while GPT creation remains more restricted than many readers may assume from older coverage. The common idea is the same across all of them: move repeated instructions out of the prompt box and into the product. The differences lie in who may build, what can be stored, and how much of the assistant’s behaviour survives from one conversation to the next.