Tier 04 / Intermediate
Every LLM ships a layer above plain chat where you can save standing instructions, attach reference files, and make a workspace specifically for one kind of work. ChatGPT calls it Custom GPTs and Projects. Claude calls it Projects, Skills, and Artifacts. Gemini calls it Gems. The names differ. The idea is the same.
01 · The big idea
A plain chat starts from zero. You explain who you are, what your project is, what voice you want, and what good looks like. Every. Single. Time. A workspace remembers all that for you. Once you set one up, the prompt you write inside it becomes one-line short.
Your role, your audience, your style, your rules. Written once, applied automatically to every conversation inside this workspace.
Documents, brand guides, contracts, examples. The model can read them as background knowledge for every chat without you re-uploading.
Conversation history that lives inside this workspace, not your global account. Old threads stay relevant to the right project.
One setup, used for months. The investment compounds. Most beginners never make it.
If you only do one thing from this tier: pick one repeatable task you do every week and build a workspace for it. The first time you use it, the savings are small. By the tenth time, you'll wonder how you ever worked without it.
02 · Three names. One idea.
Same capability, different vocabulary, slightly different feature surface. Internalize this and you'll stop being lost when someone says "build a Gem for that."
| Capability | ChatGPT | Claude | Gemini |
|---|---|---|---|
| Standalone reusable assistant | Custom GPT | Project (with system prompt) | Gem |
| Folder for related chats + files | Project | Project | (part of Gem) |
| Cross-conversation memory | Memory (global) | Memory (project-scoped, opt-in) | Memory (global, opt-in) |
| Live-rendered code/doc panel | Canvas | Artifacts | Canvas |
| Persistent instructions | Custom instructions / GPT instructions | Project instructions / "system prompt" | Gem instructions |
| File knowledge base | Custom GPT files (up to ~20) | Project knowledge (up to a window) | Gem knowledge (Drive files) |
| Web search | ChatGPT Search | Web search (toggleable) | Grounding with Google Search |
| Tools/connectors | Actions, MCP connectors, third-party apps | MCP connectors, Skills | Workspace apps (Gmail/Docs/Drive natively) |
03 · ChatGPT intermediate
ChatGPT can remember details across conversations, your name, your job, projects you're working on, preferences. Turned on by default for paid users in most regions.
Things you've told it that look "biographical" or "preference-shaped." It will literally say "I'll remember that" when it does.
Settings → Personalization → Memory. You can browse, delete, and turn it off.
Memory is global. Anything it remembers shows up in every chat. Don't let it remember details from one client that should never bleed into another.
A folder. Lives in your sidebar. Holds related chats and uploaded files. You set custom instructions that apply to every chat inside it.
"My business," "Q3 marketing," "renewing the lease." One project per recurring stream of work.
Drop in the documents that matter (style guide, contract, examples). The model can read them in every chat inside the project.
"You are helping me with X. Tone is Y. Always do Z." Set once.
If you start it in the main chat instead, none of this applies. The discipline is the whole point.
A reusable specialized assistant. You define a name, a description, instructions, knowledge files, and optionally "actions" that let it call external APIs.
The GPT Store has thousands. Search before you build from scratch. Someone has probably built 80% of what you need.
The GPT Builder is a chat-based wizard. You describe what you want, it drafts the instructions, you refine.
Custom GPTs can be private (just you), shared via link, or public in the store. Start private.
Included in Plus. Just build it.
Project = a folder for your related chats. Custom GPT = a standalone tool you (or others) launch from outside any folder. Use a project when the work is yours and recurring. Build a Custom GPT when you'd want to use it from anywhere, or share it.
04 · Claude intermediate
Claude Projects are folders with their own system prompt and knowledge files. Same idea as ChatGPT Projects, slightly different shape.
Sidebar → Projects → New. Name it, give it a short description.
Sometimes labeled "system prompt." Write the role, audience, voice, constraints. Be specific. This is your one-time investment.
Drop in PDFs, docs, text files. Claude's long context means you can fit a lot. The project will use them in every chat inside it.
Claude memory is opt-in and project-scoped, it learns from chats inside this project, not globally. Cleaner separation than ChatGPT's global memory.
A side panel that renders whatever Claude is generating, code, a document, a chart, a small interactive app. You can iterate on it live without copy-pasting back and forth.
Automatically when the output is "thing-shaped", a document, code, an SVG, a small webpage. You can also explicitly ask.
Edit in chat, download, copy, publish as a small standalone site (for shareable artifacts).
Cuts the "copy → paste → edit → repaste" loop. The artifact is the document; the chat is the editor.
Claude's reusable instruction packages. Less marketed than Custom GPTs but increasingly central. A Skill is a folder containing a markdown file (SKILL.md) with instructions Claude loads when the description matches your task.
Repeatable, structured workflows. "Always do X this way." "When asked to do Y, follow these steps."
By description match, not name. You don't summon a skill; Claude loads it when the work fits.
Inside a Project, or globally (Pro+). Built and shared as folders. More technical than Gems or Custom GPTs but more powerful.
Skills are the most powerful intermediate-feature in any of the three LLMs, but they have a learning curve. As a beginner, start with Projects. Come back to Skills when you have a workflow you've repeated five or six times and want it codified.
05 · Gemini intermediate
Gemini's version of a Custom GPT. Reusable specialized assistants with their own instructions and knowledge.
Google ships several useful ones: writing editor, career coach, brainstormer. Try them; some are surprisingly good.
Gem manager → New Gem. Write the instructions, attach Drive files as knowledge, save.
Otherwise they get lost. Two or three live Gems is the right number.
Gemini's superpower nobody else has. It lives inside Gmail, Docs, Sheets, Slides, Drive, Meet, Calendar. With AI Pro turned on, every Workspace app gets a side panel that knows your data.
"Summarize this thread." "Draft a reply that asks for X." Reads the thread context automatically.
"Generate a first draft." "Pull insights from this data." "Rewrite this paragraph tighter." Lives in the sidebar.
"Find the contract with the indemnity language." Searches across your files semantically, not by filename.
Live transcripts, action items, summaries after the call. Works automatically if turned on.
A separate Google product, technically, but if you're a Gemini user it's worth a tab. Drop in your sources (PDFs, articles, websites, YouTube videos), and it builds a grounded notebook that can answer questions, generate study guides, and produce audio overviews.
Anchoring AI answers to a specific, finite set of sources. Almost no hallucination, every answer cites which source it came from.
"Audio Overview" generates a podcast-style conversation about your sources. Sometimes uncanny, often surprisingly good for learning.
Anytime you have a set of documents you want to actually understand, not just summarize. Research, study, due diligence.
06 · Voice mode
All three LLMs have a voice mode. Most users open it once, find it weird, never come back. That's a mistake, voice is the single biggest unlock for thinking with AI, not just typing at it.
| ChatGPT Advanced Voice | Claude Voice | Gemini Live | |
|---|---|---|---|
| Where it shines | Natural conversation, role-play, language practice | Long-form thinking, careful answers | Camera + voice on Android; pointing at things |
| Latency | Near-real-time, can interrupt | Slightly slower, more turn-based | Near-real-time, can interrupt |
| Best surface | iOS app | Mobile app | Android (Pixel especially) |
| Underrated use | Walking-and-thinking dictation: dump 10 minutes of messy thoughts, get a structured summary | Read me back this article and quiz me on it | Point camera at a board, a chart, a UI, a recipe; talk about it |
The unlock isn't "AI that talks back." It's that you can think out loud and have something on the other end that actually captures it and structures it. Walk around the block, dump a ten-minute monologue, end with "now write that up as a one-page memo." You will never type that fast again.
07 · Image generation (briefly)
Image gen lives in Tier 5 as a fuller topic. At this tier you just need to know: ChatGPT and Gemini will draw for you in the same chat box. Claude won't.
08 · Worked example
A useful drill. Build a "weekly newsletter writer" in all three. Same brief. See how each shapes up.
Role: You are a newsletter editor writing one weekly issue for [audience].
Voice: Conversational, opinionated, no corporate-speak. One main idea per issue. Always include "Why this matters" and "What to do this week."
Format: Markdown. Subject line + 300-500 word body + sign-off.
Knowledge: Three example past issues (uploaded). The newsletter's positioning doc.
Constraints: No em-dashes. No "in today's fast-paced world." No bullet lists in the body unless they're load-bearing.
GPT Builder → describe the brief in chat → it drafts instructions → refine → upload your past issues as knowledge → save → pin to sidebar. Use it by typing "Topic: X" each week.
Projects → New → paste the brief into custom instructions → upload past issues + positioning doc → start new chats inside the project. Bonus: use Artifacts to draft the issue in a side panel and edit it inline.
Gem manager → New Gem → paste the brief → attach past issues from Drive as knowledge → save → pin. Use it weekly. Bonus: ask Gemini to pull research from your Gmail/Drive on this week's topic before drafting.
Ask each one to draft the same week's issue. Read all three. You'll notice immediately: Claude's prose is the most natural, ChatGPT's structure is the cleanest, Gemini will likely pull in something from your actual Drive that the others can't. None is "best." All are useful for different reasons.
09 · When to skip all this
A common Tier 4 mistake: building a Custom GPT for every one-off. The setup cost only pays back if you're going to use it ten times.
When the task repeats. When the standing context is heavy. When you'd want to share the setup with a teammate. When you'll forget your instructions a week from now.
When you're exploring. When the task is genuinely one-off. When you don't yet know what "good" looks like. When you're learning the shape of the problem.
10 · Before you climb