Working with AI · 12 min read
I’ve been running my entire marketing consultancy on Claude since September 2023. One tool. 25 active Projects. Here’s how the setup works and what I give up for it.
NA
Natalie AndersonMarketing consultant · Farmingdale, Maine · October 7, 2026
I canceled my ChatGPT subscription after it told me “you’re not crazy” during a troubleshooting session. Unprovoked. Nobody asked.
I was already frustrated. The sycophancy had been building for months. Every response felt like it was managing my feelings instead of answering my question. But that line crossed something for me. I don’t need a chatbot to validate my sanity.
Then the Pentagon situation happened. Anthropic pushed back on a Department of Defense arrangement it wasn’t comfortable with. OpenAI stepped in and took the deal. I’m not saying Anthropic is perfect. They’ve made questionable decisions too. But when your competitor picks up the contract you refused on principle, that tells me something about priorities. I stopped using ChatGPT that week. Haven’t opened it since.
I’ve been running my entire marketing consultancy on Claude since September 2023. One tool. 25 active Projects. Client work, my own business, content, web development, personal stuff. All of it.
I’m not here to argue that Claude is better than ChatGPT. What I can tell you is what happened when I stopped splitting my attention across five AI tools and committed to one. The architecture I’ve built over three years would work with any AI tool that supports persistent context. I just happen to believe I picked the right one.
1.Stop Splitting Your Work Across Tools
I see people using ChatGPT for writing, Gemini for research, Midjourney for images, Perplexity for search, and Claude for “the hard stuff.” They spend half their time copying context between platforms and the other half re-explaining who they are.
Every tool you add is a relationship you have to maintain from scratch. None of them know about the others. None of them remember what you told the last one. You are the integration layer, and you are doing it manually, every single day.
When I committed to Claude, something shifted. Instead of teaching five tools the basics, I taught one tool everything. My voice. My clients. My standards. My bad habits. Three years of accumulated context that no new tool could match even if it were technically superior.
The setup · 1 tool · 25 Projects · since Sept 2023
“Instead of teaching five tools the basics, I taught one tool everything. My voice. My clients. My standards. My bad habits.”
A new client project takes about an hour to set up now because the structure already exists. That hour is not something you want to rush.
The compounding effect is real. The longer you work with one tool, the less setup each new task requires. A new client project takes me about an hour to set up now because the structure already exists. It isn’t something you want to rush, but it also isn’t a week of onboarding.
There’s also something that’s harder to quantify. Claude has a memory feature that synthesizes your conversations and learns your preferences, your working style, your project context. Anthropic’s own framing is that “each conversation improves the next.” When I moved a client from my main workspace to their own Project, the output quality dipped for a while. Part of that was me learning how to structure a Project well. But part of it was starting fresh without the accumulated context Claude had built from months of working together. It took time to get back to the same level.
I’ve seen some debate about whether this is real or just perception. Based on three years of daily use, it’s real. The mechanisms are documented: memory, project knowledge files, conversation history that Claude can search. The model isn’t literally training on your data. But the accumulated context around it creates something that functions like institutional knowledge. A new Claude conversation inside a well-built Project with months of memory behind it performs meaningfully differently than a cold start.
2.Set Up a Project for Every Client, Venture, and Area of Your Work
Claude has a feature called Projects. Think of each one as a dedicated workspace with its own memory, reference files, and instructions. Everything you put in a Project stays available to every conversation inside it.
I have 25 active Projects right now. That sounds like a lot until you see how they break down.
Current clients
One Project each.
My business
Natalie Anderson Creative has its own.
New venture
A couple for the business my husband and I are starting.
Past clients
Still have theirs, in case I pick the work back up.
Personal
Meal planning, household stuff.
The key is that each Project is self-contained. When I open a client’s Project and start a new chat, Claude already knows their brand voice, their content history, the tools they use, and how they like things structured. I don’t re-explain any of it. I just start working.
If you’re using Claude without Projects, you’re essentially introducing yourself on every first date and wondering why the relationship never deepens. Set up a Project for every area of your work. It takes ten minutes and it changes everything about how you use the tool.
3.Build Your Project Knowledge Files (This Is Where the Real Advantage Lives)
A Project without knowledge files is just a folder. The files are what make it intelligent.
For every client Project, I include some version of the same core documents. A voice guide that captures how that client sounds, built from real writing samples. My AI tells document, which is a list of patterns AI produces that weaken writing (mirror constructions, stacked lists, throat-clearing openers, and others I’ve cataloged over three years of daily use). An overused words list so Claude avoids the vocabulary that screams “a robot wrote this.” A brand guide if the client has one. Samples of past newsletters, social media posts, or whatever content types I produce for them.
For client projects that involve a specific initiative, I also include a detailed SOP for that project. If I’m building a website, the Project has the site architecture and design specs. If I’m running a content calendar, it has the content pillars and posting strategy.
The non-client Projects look different. My personal NAC Project has my own voice rules, my content strategy, my LinkedIn posting framework, and my business SOPs. A project for a side venture might just have a product brief and some market research.
The pattern is the same even when the contents vary. Every Project gets the reference material Claude needs to do the work without being told how, every single time.
If you’re starting from scratch today, these are the files I’d create first.
1
Voice samples
Two or three pieces of writing that sound like you (or your client). Not your best work. Your most representative work.
2
An overused words list
Start with the obvious ones: “crucial,” “game-changer,” “compelling,” “resonate.” Add to it every time you catch Claude defaulting to a word that makes you cringe. These aren’t banned words. You’ll use some of them, and that’s fine. The point is to notice when your AI reaches for the same ones repeatedly. When it does, a thesaurus is your best friend. I put together a starter list you can copy and add to: AI Overused Words List.
3
Brand basics
Colors, fonts, tone descriptors. Even a few bullet points help.
4
A "don't do this" file
AI patterns you’ve noticed and want eliminated. This is the single most impactful file you can create, and I wrote a whole article on how to build one: 5 AI Writing Patterns That Are Ruining Your Content.
5
Context about the work
What are you producing? For whom? What does success look like? Two paragraphs is enough to start.
4.Start Fresh Chats With a Summary
Long conversations degrade. Around 30 to 40 messages, Claude starts losing the thread. It forgets what you decided earlier. It contradicts itself. It gets repetitive. This isn’t a flaw you can power through. It’s how the technology works right now.
More recently, Claude has started compacting conversations earlier than it used to, condensing the context mid-chat in ways that can stall your workflow. You’ll be in the middle of something and the compaction kicks in, loses a detail you need, and now you’re spending time re-establishing context instead of working. I’m still figuring out the best way to handle it (and honestly hoping it’s a bug that gets addressed), but it makes getting ahead of long chats even more worthwhile. Shorter, focused conversations with good summaries between them avoid the problem entirely.
My fix is simple. When a conversation starts getting long or losing coherence:
1
Ask Claude for a summary
Everything we’ve discussed, including decisions made, open questions, and next steps.
2
Start a fresh chat inside the same Project
Same knowledge files, same memory, clean slate.
3
Paste the summary at the top
Reference the chat you moved from so Claude can search back to it if needed.
4
Keep going
Takes 30 seconds. I do this multiple times a week, sometimes multiple times a day.
The new conversation picks up with full clarity because it has the summary as context plus all the Project knowledge files. It’s like handing a detailed briefing document to a sharp colleague on their first day. They don’t need to have been in the room for every conversation. They need to know what was decided and what’s next.
It keeps the quality of Claude’s output from sliding.
5.Use Conversation Search to Find What You’ve Already Done
This is the feature most people don’t know exists, and it changed how I work.
Claude can search through your past conversations. Not just the current chat. All of them, within a Project. You can ask it to find the LinkedIn post you drafted two months ago, pull up the client onboarding notes from last quarter, or locate that one conversation where you worked out the pricing for a new service.
I use this constantly.
Before a blog post
Search for previous conversations where we discussed the topic.
Before a client call
Pull up our last few working sessions so I’m caught up.
When I’ve lost a decision
Search for where I landed instead of recreating the thinking from scratch.
The practical effect is that three years of working conversations become a searchable knowledge base. Every draft, every brainstorm, every strategic decision I’ve worked through with Claude is retrievable. That accumulated history has real value, and I genuinely cannot overstate how much it changes the way you work. It’s like having a brilliant assistant who can recall that newsletter restructure you discussed six months ago but never implemented and can’t quite remember the details of. The idea is still there, fully intact, in a conversation you can pull up in seconds. That only exists because I’ve been working inside one tool the whole time.
If I’d been splitting my work across ChatGPT, Gemini, and Claude, that institutional memory would be fragmented across three platforms. None of them would have the full picture.
6.Pick Your Model by the Task
Claude isn’t one model. It’s a family of them, and they’re not interchangeable.
For writing, I use Opus 4.6. It’s the most careful with voice, the least likely to steamroll my phrasing, and the best at taking a light editorial hand instead of rewriting everything it touches. Newer models have launched since Opus, but for writing specifically, I haven’t found anything that matches it.
For general tasks like updating an Airtable, organizing information, or running a browser extension, Sonnet works well and uses less of my daily quota. It’s fast and capable for anything that doesn’t require a nuanced writing voice.
For web development work through Claude or Claude Code, I use Fable. It handles code, WordPress modifications, and technical builds effectively. I would never use it for writing. It’s verbose and takes over the process in ways that Opus doesn’t.
Writing
Opus 4.6
Client newsletters, blog posts, anything with a voice.
General tasks
Sonnet
Airtable updates, organizing files, browser extension work. Lighter on quota.
Web development
Fable
Code, WordPress modifications, landing pages. Never writing.
The trick
Start in Fable to set up the framework, then switch to Sonnet to do the actual work. Fable burns through quota fast.
One trick I’ve landed on: I’ll sometimes start a task in Fable to get the framework set up, the outline and the initial direction. Then I switch to Sonnet to do the actual work. Fable burns through quota fast, so using it just for the scaffolding and handing off to Sonnet for execution saves a significant amount of daily usage.
The point is that “using Claude” doesn’t mean using one thing. It means understanding which version of the tool fits the work in front of you. I switch models mid-day depending on what I’m doing. Writing a client newsletter? Opus. Building a landing page? Fable. Updating a spreadsheet or organizing files? Sonnet. Same platform, matched to the task.
7.What I Give Up by Using One Tool
I should be honest about the trade-offs because this isn’t a “Claude is perfect” article.
Image generation
Image generation is the obvious one. Claude has started adding visual capabilities (Claude Design creates programmatic graphics and layouts), but it still doesn’t do traditional image generation the way Gemini or DALL-E do. For actual images, I still rely on Gemini and Gamma, which technically breaks my one-tool rule. But image generation is a discrete task with no context dependency. I don’t need the image generator to know my voice or my clients.
Usage limits
Claude also has usage limits. On busy days, I hit the cap and have to wait or switch to a lighter model. This is genuinely annoying when I’m in the middle of something time-sensitive.
Vendor lock-in
The biggest real trade-off is vendor lock-in. Three years of context, workflows, and project architecture live inside one platform. If Claude’s quality degraded significantly, or if Anthropic made decisions I couldn’t support, migrating would be painful. I’m aware of this. I’ve accepted it because the alternative, spreading everything thin across multiple tools, has a higher daily cost than the theoretical risk of needing to move. That said, I’m in the early stages of building a contingency plan. When you depend on any single tool this heavily, it would be foolish not to have an alternative ready in case of a major outage, a buyout, or just a shift in my comfort level with the company’s policies. The same values-based decision that brought me to Claude could take me somewhere else.
It gets things wrong
And sometimes Claude is just wrong. Every model is. They hallucinate. They confidently suggest the easy way when the right way is harder. Every AI tool does this. You have to read the output. You have to understand the content well enough to spot when something is off. And when you’re not sure, you have to ask, push back, verify. The tool doesn’t replace your judgment. It extends your reach. But only if you’re actually paying attention to what it gives you.
Free starter file
The AI Overused Words List
The vocabulary list I load into every Project so Claude stops reaching for “crucial” and “game-changer.” Copy it, then add your own cringe words as you catch them.
Opens a Google Doc you can copy into your own Drive.
Tools I Recommend
Tools I use and have vetted. Not all were used for this specific article.
Claude
My only AI tool since 2023. Projects hold the voice rules and reference files that make this workflow possible.
Descript
Audio and video editing for podcast production. Where raw recordings become finished episodes.
Riverside
Remote recording that captures each side locally. Solves the audio quality problem before editing starts.
Gamma
Turns reference docs into designed PDFs and presentations without opening a design tool.
Disclosure: Some links are affiliate links. I’ve used and vetted every tool I recommend. It costs you nothing and does not affect what I recommend.
Frequently Asked Questions
Do I need a paid Claude plan for this?
Yes. The free tier doesn’t include Projects, which is the foundation of everything described here. The Pro plan ($20/month) unlocks Projects and gives you significantly more usage. There’s also a Max plan ($100/month for 5x usage, $200/month for 20x) if you need more capacity. I’m on Pro and rarely hit the limit now that I’ve learned to use lighter models like Fable for scaffolding and Sonnet for general tasks, saving Opus for writing. I upgraded to Max for a stretch, but learning how to manage model usage let me step back down without losing anything.
How long does it take to set this up?
Your first Project with basic knowledge files takes 30 minutes to an hour. Once you have reusable files saved and accessible (like your AI tells document and overused words list), new Projects take about an hour, and that hour is worth protecting. Building out a full voice guide and AI tells document takes longer, but you can start simple and add files as you learn what helps.
Can I use this approach with ChatGPT or another AI tool?
The principles apply anywhere you can create persistent context. ChatGPT has Custom GPTs. Gemini has Gems. The specific features I describe (Projects, conversation search, model switching) are Claude’s implementation, but the strategy of committing to one tool and building institutional knowledge works regardless of which tool you choose.
What if Claude's quality drops or Anthropic does something I disagree with?
That’s a real risk, and I take it seriously. I’m actively building a contingency plan for exactly that scenario. The daily cost of fragmented tools is still higher than the theoretical cost of migrating, but it would be foolish to depend this heavily on any single platform without an alternative ready. Your calculus might be different.
Is this only useful for marketing work?
No. The architecture works for any knowledge work. I use it for marketing, content, web development, business strategy, and personal organization. The Project structure adapts to whatever you do.
Free reference guide
The AI Writing Tells Checklist
All 14 patterns with before-and-after examples and fixes. Drop it into your Claude project or custom instructions so your AI knows what to avoid. Subscribers also get new posts on working with AI as they publish.
One useful email at a time. Unsubscribe anytime.