We built 3,000+ agents in ChatGPT before discovering why Claude was wasting 63% of tokens on filler. Here’s everything we learned in almost 3 years.
Hub365 has been using Claude since July 2023. Before that, we built 3,000+ custom agents in ChatGPT. Almost three years later, we learned something no tutorial teaches: AI doesn’t fail because it’s limited. It fails because you set it up wrong. This blog shares our real timeline, the discovery that cut token waste by 63%, and how Claude Projects transformed what we deliver to clients. Includes a free token calculator and a prompt builder that generates your custom instructions in under 2 minutes.
I’m going to be honest with you.
The first time I used Claude, I thought it was worse than ChatGPT. Slower, more cautious, less “impressive.” ChatGPT gave you answers that sounded brilliant in 3 seconds flat. Claude took its time and sometimes told you “I’m not sure about this.”
That was July 2023. The exact day Anthropic launched Claude 2 to the public. Todd and I signed up that same afternoon. We’d already been using ChatGPT for 8 months, had built over 3,000 custom agents (yes, three thousand — Todd says it was like raising 3,000 kids who all talk at the same time and none of them listen), so we had a solid point of comparison.
Today, almost three years later, Claude is the backbone of what we deliver to our clients. Marketing strategies, bilingual content, automated follow-up sequences, visibility audits, conversion funnels, and complete digital ecosystems for hundreds of clients across 12+ industries. Everything runs through Claude. ChatGPT is still in our toolkit, but for different tasks. And figuring out which tool belongs where took us months of trial and error.
This isn’t a generic Claude tutorial pulled from the internet. It’s the real story of how a bilingual marketing agency built its entire operation on this platform, with dates, mistakes, and the lessons we wish we’d had from day one.
Our Timeline with Claude: From Day 1
November 2022
ChatGPT enters the scene. We start experimenting. Over the following months we build over 3,000 custom agents. But somewhere between agent 500 and agent 2,000, we notice a pattern: ChatGPT was excellent at generating volume but wildly inconsistent on quality. Statistics that didn’t exist. Sources fabricated with complete confidence.
March 14, 2023
Anthropic launches Claude 1. Limited access only. While everyone else competed to be fastest, Anthropic competed to be most honest. That got our attention.
July 2023
Claude 2 opens to the public. We sign up the same day. Claude didn’t make things up. If it didn’t know something, it said so. For someone who builds content with their name behind it, that’s not a limitation — that’s a guarantee. Same thing we demand from a website that represents your business: honesty first, impress later.
October 2023
The “why does Claude get tired?” moment. In longer conversations, Claude started giving generic answers, losing the thread. I told Todd, “Claude is getting tired.” It wasn’t fatigue — it was context overload. Every message dragged the entire conversation history plus thousands of words of instructions. That was our first real breakthrough: the problem wasn’t the AI. It was our architecture.
March 4, 2024
Claude 3 changes the game. Opus, Sonnet, Haiku. Three models, three capability tiers. Right tool for the right job — a principle we also preach when talking about the 5 tools every business needs.
June 2024
Claude Projects and Artifacts launch. Persistent workspaces where you upload documents and define instructions Claude remembers across every conversation. We loaded brand manuals, SEO strategy guides, product catalogs, and automation flows. Claude went from chatbot to team member.
February 2025
Claude Code arrives as preview. Todd nearly cried. An AI agent working directly in the terminal, reading code, executing tasks, making real changes to files. For an agency managing WordPress, n8n, Make, Zapier, and multiple integrations to automate client follow-up, this was next level.
May 2025
Claude 4 and Claude Code go GA. Full-scale production. Complete bilingual content, technical specs, interactive design, smart chatbots plugged into the CRM. All from one platform.
January 2026
Claude Cowork launches. Claude Code but for non-developers. Works directly with local files, keeps context across sessions, automates repetitive desktop tasks. Just like DeepSeek democratized access to advanced AI, Cowork democratized AI productivity for those of us who don’t write code.
February 5, 2026
Opus 4.6 with 1 million token context. Context window jumped from 200,000 to 1 million tokens. Clients in healthcare, real estate, funeral services, and luxury industries felt the difference immediately. What used to take five sessions now resolves in one. The kind of infrastructure that separates a CRM that actually captures leads from one that just takes up space.
March 2026
Cowork gets Projects. Persistent workspace functionality now connected directly to your local files. Full circle.
ChatGPT vs Claude: What We Learned After 3,000+ Agents and Almost 3 Years
After building over 3,000 custom agents in ChatGPT and migrating our core production to Claude, we have something few people can claim: real experience, at scale, with both platforms for almost 3 years.
We use both. Every single day.
ChatGPT wins at:
Fast idea generation, brainstorming, first drafts when you need volume, exploratory research where speed matters more than accuracy, and structuring what you want before you build it.
Claude wins at:
Final content production with your brand voice, work that needs consistency across long documents, any task where accuracy matters more than speed, following complex instructions without going off the rails. It’s the difference between being found by Google and being recommended by AI: one thing is showing up, another is being cited with confidence.
The problem with ChatGPT for production: it tends to tell you what you want to hear. Claude tells you “there’s a problem in paragraph three” and explains why. For an agency whose reputation is built on honesty, that difference is worth more than any new feature.
The Architecture Mistake That Was Costing Us Thousands of Tokens
When Claude Projects launched in June 2024, we made the same mistake everyone makes: we dumped everything into Instructions. Brand guide, tone of voice, service descriptions, client personas, do’s and don’ts, formatting rules, SEO guidelines. All of it. We went from a 400-word instructions file to over 3,000 words.
Result: Claude got slower, more generic, and started contradicting itself. More instructions were producing worse outputs. We were confused.
Todd ran the numbers:
The breakthrough came when we separated two things we’d been treating as one: Instructions vs Knowledge.
The Framework
Instructions = behavior rules only. How Claude should think, communicate, prioritize. Maximum 500 words. No exceptions.
Knowledge = everything else. Brand guide, service descriptions, client personas, SEO frameworks, formatting examples. Uploaded as separate documents, searched on demand.
Instructions load with every message (constant token cost). Knowledge is pulled only when referenced (cost only when used).
Same logic we apply when explaining to clients why their business needs digital structure, not patches.
The Reliability Protocol
Efficiency without accuracy is dangerous. Same standard we demand from a LinkedIn profile representing a business: every word holds up on its own.
Hub365 Reliability Clause: “Hub365’s reputation is built on honesty and trust. Every output must be verifiable. Zero tolerance for fabricated stats, invented sources, or unverified claims. If you are unsure, say so.”
We added this to every single project. And the output quality went up — not because Claude was suddenly “smarter,” but because we’d made the rules explicit. Like the way Gemini and ChatGPT handle scheduled actions better when given structured parameters: the model doesn’t guess what you want.
Token Calculator: How Much Are You Wasting?
Move the sliders, plug in your real numbers, and see how much you’re throwing away every month.
6 Things You Can Fix Right Now
Move brand guide to Knowledge
Instructions = behavior. Knowledge = reference. Each in its place — same logic behind the 5 tools every business needs.
Add the 8 efficiency rules
"Answer first. No preamble. No flattery." Shorter, better responses immediately.
Add a reliability clause
"Never fabricate stats or sources. If unsure, say so." One sentence. Massive impact.
Name reference documents explicitly
"Before producing content, search Knowledge for [Document Name]." Claude won't assume. Tell it.
Keep Instructions under 500 words
If you can't say it in 500, it's reference, not a rule. Move it to Knowledge.
Use ChatGPT to think, Claude to produce
3 monitors: ChatGPT for exploration, Claude for production, CRM on the third. No switching windows. No losing the thread.
AI Changes Every Day. Your System Doesn't Have To.
In almost 3 years: Claude 1, 2, 3, 3.5, 4, 4.5, 4.6. Projects, Artifacts, Code, Cowork. Every update brought the temptation to start over. But the base survived every version.
Build Your Custom Claude Project Prompt
6 questions. 2 minutes. Ready-to-paste Instructions file.
Or comment CLAUDE on our social posts for DM delivery
Frequently Asked Questions
A persistent workspace that Anthropic launched in June 2024 where you store custom instructions and reference documents. Everything inside is available to Claude across every conversation without repeating context.
Every token in your Instructions gets processed with every message. If your instructions are 4,000 tokens, that's consumed before Claude reads your actual question. Over 20 messages that's 80,000 tokens just re-reading instructions. That's why ours clock in at 387 tokens and nothing more.
Projects require Pro ($20/month) or above. The free plan doesn't include Projects. However, the behavior rules can be pasted at the start of any conversation on any plan. You lose persistence but gain efficiency.
Instructions load on every message (constant cost). Knowledge gets searched on demand (cost only when pulled in). It's the difference between paying rent year-round and paying only when you use the space.
Claude Code is terminal-based, built for developers. Cowork is the visual interface version that Anthropic launched in January 2026 as a research preview, designed for people who don't live in code. Requires Claude Max ($100–$200/month) and is currently macOS only.
We use both for different things. ChatGPT for rapid exploration and brainstorming. Claude for final production, branded content, and any work where accuracy and consistency matter more than speed. After almost 3 years and over 3,000 agents built across both platforms, the combination wins. Not a single model.
Your AI Is Only as Smart as Your Instructions
Comment CLAUDE on any of our social posts and we’ll DM you the prompt builder link.
Or get the full ecosystem: AI chatbots, SEO + GEO, and a website that converts.
