I’m pretty sure everyone has tried using an LLM to create content at least once by now. Many have even figured out how to get content that doesn’t suck.
I was able to produce some pretty good articles last year, but the effort involved meant I didn’t really want to do it very often.
That’s the biggest issue I’ve had with AI in the past couple of years. It can do some pretty amazing things, but getting it to an acceptable level is such a pain, and requires a lot of agency.
I’d even argue that LLMs made it harder to create good content. Most people aren’t happy with their output, but writing things the old way just feels so archaic, I might as well have been writing with ink and quill.
I tried creating a customGPT and using Claude projects. They help to some extent, but are a pain to set up and require constantly changing prompts and updating information.
Then every time I used them I was staring at a blank page trying to one-shot an article or preparing to tweak it until ChatGPT forgot what I was talking about.
That is, until Claude Code and Opus 4.5 changed everything.
I’ve produced more content in Q1 than I produced in the previous 4 years, and if I do say so myself, the content is actually good. It gets better the more I interact with Claude, and it’s a joy to use.
Claude not only writes well, it uploads and schedules my content (with the occasional mistake that I’ve since rectified).
The real unlock with Claude and Opus has been ease of use. Yes there is a learning curve, but I’m not talking about some fancy n8n automation that only the nerds know how to use. I’m talking about sitting down and saying “alright let’s work on the next batch” and it knows what I’m talking about, pulls up a list of ideas and drafts in progress, and gets to work.
It knows my voice. It remembers my rules. It produces multiple formats at once and distributes them across every channel I care about. And it gets better every week because my feedback becomes permanent.
The closest I’ve ever had to this was back in 2023 when I was paying $8,000 a month for a content agency that still couldn’t get my voice quite right, and I was spending ~3 hours per week approving content, coming up with ideas, and doing work I didn’t want to.
Now though, I publish across five channels weekly, the output is better than what the agency produced, and the system costs a fraction of what I was paying, with less time commitment once it’s all set up.
In case you are thinking “I have a customGPT already”, I want to explain why you can build something so much better now.
In fact, I asked Claude to explain it for me.
What a CustomGPT Does
A customGPT is a chatbot with a system prompt. You type “write me a LinkedIn post about our Q3 results” and it generates text. Every time. From scratch. The “customization” is a static instruction block that tells it tone, audience, and maybe some company facts.
That’s it. That’s the whole system.
What a Content System Does (That a CustomGPT Cannot)
1. It knows your full strategic context, not just voice guidelines
A content system has your audience segments, your content pillars with rotation tracking, your funnel stages, and your positioning baked in. A customGPT might know “write in a casual tone.” A content system knows why you’re writing, who it’s for, and where this piece fits in the larger narrative arc.
Every piece gets tagged with a pillar, an audience segment, and a funnel stage. That means you can look back and say “we’ve been heavy on product content for existing customers, time for a thought leadership piece aimed at new prospects.” A customGPT has no concept of what you wrote last week.
2. It produces a full distribution package, not a single output
One input becomes five formats (newsletter, X thread, LinkedIn post, video script, blog post) plus branded images in two sizes, scheduled across multiple channels. A customGPT gives you one blob of text you then manually reformat and paste everywhere.
3. It actually publishes
The system calls APIs to schedule social posts, create scheduled blog posts with featured images and categories, and set up email campaigns. A customGPT can’t touch any external system. It just generates text you copy-paste.
4. Guardrails are structural, not suggestions
My system has hard rules that shape every draft. Your business will have its own: brand guidelines, legal requirements, tone rules, things you never want to say. These aren’t “please remember to stay on brand” instructions. They’re woven into the drafting process and enforced every time. A customGPT might have a line saying “follow brand guidelines” but has no mechanism to enforce it consistently.
5. It has a content calendar and memory of what’s been published
A content system tracks what’s been published, what’s coming up, and the narrative arc across weeks. It reads this at session start. A customGPT starts every conversation at zero.
6. It has built-in workflows
- A weekly prompt that checks recent feedback, pillar usage, and suggests ideas
- Templates that extract content angles from meetings, calls, or interviews
- A lead magnet pipeline with end-to-end build automation (PDF, landing page, forms, email sequences)
- Company account reshare workflows that post to your brand pages after your personal posts go live
A customGPT is a blank text box every time you open it.
7. It improves itself
A feedback file captures reactions and observations. At each session start, the system reads it, summarizes, and asks if there’s anything new. If a pattern is durable, it gets promoted to a permanent rule. If it’s one-time, it gets cleared. The system’s instructions evolve. A customGPT’s instructions only change when someone manually edits them.
8. It has a filesystem, not just a prompt
Drafts, scheduled posts, published archive, images, planning docs, reference materials, scripts. The content system is a structured workspace with dozens of files across organized directories. A customGPT has a text box and maybe some uploaded PDFs it can search.
The Shortest Version
A customGPT is a copywriter who shows up every day with amnesia and writes one thing at a time.
A content system is an integrated content operation: strategy, production, distribution, scheduling, publishing, tracking, and self-improvement, that happens to use AI as the engine.
It’s like comparing Google Docs to an ERP.
Now, it does take a few weeks to set up and get working well, but it pays dividends after that.
Here’s how to build the same thing for yourself.
Side note: You can also hire Onfolio to build and manage this system for you, and it costs a lot less than $8,000/mo for better results and almost unlimited quantity.
Step 1: Getting Claude to help define your content pillars and audience
This is basic content marketing, but what used to be a boring chore is now a 20 minute job. You need to answer two questions. Who are you writing for? And what do you have the authority and experience to write about?
Most founders skip this and jump straight to “write me a LinkedIn post about leadership.” That’s how you end up sounding like every other numpty on the platform.
The point of this guide is teaching you how to build an AI system. I shouldn’t spend much time on the basics of content marketing.
But how did I come up with my pillars? Claude helped me, obviously.
I asked Claude to use the QuestionTool to interview me about what matters to me, what matters to my audience, who the various members of my audience are, and let it do the hard work.
Step 2: Document your voice
This is the step that separates a content system from ChatGPT.
You won’t need to spend more than 10 minutes doing this initially, and the Claude Code system you build will allow you to keep improving it and perfecting it as you go.
That’s really a big part of the beauty of this. It’s frictionless to keep improving the system on the fly as you use it.
Everything else I’ve tried to do with AI is such a chore I often can’t be bothered to do it.
With Claude Code it’s all so easy I’m constantly asking myself “What can I optimize next?”
A real voice document is 1-2 pages that captures how you actually communicate.
I’ve never read mine.
I had Claude interview me and I provided it with a bunch of articles I’ve written and then it created the document. I improve it by asking it to write articles for me and telling it what I do and don’t like, and it automatically updates it.
Over time, this document becomes incredibly precise. That’s the compounding effect that makes AI content systems work.
Tip: Share some of your favorite content that you’ve produced with Claude. The more the better though, or it will over-index on a sentence you wrote once and replicate it every time.
Step 3: Set up the AI layer
Claude Code is a command-line tool that reads a local instruction file every time it runs.
My instruction file is about 15 pages and covers everything: voice rules, compliance requirements (I run a public company, so this matters), content formats, distribution workflow, even which image sizes to generate for each platform. When I say “draft a newsletter about our Q3 results,” it already knows the format, the tone, the audience, the disclaimers to include, and where to schedule it.
I use Claude Code inside VSCode, which is a free wrapper that makes CC so much easier to use and is in my opinion key to the system.
You can use something like Cursor as well, but VSCode is free, and not facing the essential crisis that Cursor is (despite a $30bn valuation)
Think of it like a command centre. You have your file system on the left, your open documents in the center, and Claude Code on the right.
I can use Claude to brainstorm ideas, update documents en masse, create plans, add images, do research, or edit an article.
If I don’t like a sentence? Not only can I edit it. I can highlight it and then say to Claude something like: This sentence isn’t true. Update your memory to fix this and check you haven’t made a similar mistake elsewhere.
When I see an interesting article? I paste it into Claude and say “I want to work some of these concepts into future articles” and a new brainstorming session begins.
Claude also connects to my email software to schedule emails, schedule WordPress posts, schedule social posts, and manage an entire content calendar.
Using skills and plugins you can connect to image generation APIs and generate branded images for your articles.
Here’s a tip that will make Claude infinitely better. Add this to your custom instructions:
Unless the question is very simple, always use the ask question tool to ask me questions. Do as many rounds of questions as required until you have a clear picture of all the parameters you need to perfectly answer the question.
This single rule turns Claude from a text generator into something that actually interviews you before it writes. The output quality difference is night and day.
Step 4: A multi-format production workflow
Here’s where it gets powerful. A single input (a few bullet points, a voice memo transcript, even a rough paragraph of thoughts) should produce content for every channel you publish on.
For my system, one input generates:
- A newsletter (400-800 words, one topic, story-driven)
- A Twitter/X thread (5-10 tweets, hook-first)
- A LinkedIn post (150-300 words, strong opening line)
- A video script (60-90 seconds spoken)
- A blog post (800-1,500 words, SEO-structured with subheadings)
- Company social posts (shorter reshare versions for the company accounts)
Each format has its own structural template baked into the system. Claude doesn’t just “make it shorter for Twitter.” It knows that Twitter threads need a bold first tweet, one idea per tweet, and a CTA at the end. It knows that LinkedIn truncates after two lines so the opener needs to hook immediately. It knows the blog post needs H2 subheadings and a different CTA than the newsletter.
You build these templates once. Define the structure, length, and rules for each format. Then your production workflow becomes: provide raw input, get back all formats, review, approve, schedule.
The time savings are enormous. What used to be five separate writing sessions becomes one review session.
In the past I would get an article I liked and upload it to ChatGPT and say “make this shorter for LinkedIn” and it would either end up still too long, or it would lose the essence of the message. Most of the time I didn’t bother.
Now I don’t even have to do anything because Claude schedules it once I approve.
Step 5: Set up distribution and scheduling
Speaking of which, content that sits in a Google Doc doesn’t do anything. You need a way to schedule and distribute across every channel automatically, or as close to automatically as possible.
Scheduling and formatting content has always been a massive pain for me. I hate the monotony even if it’s just a quick copy paste and add a few formatting tweaks.
It is beneath me.
Now I type “schedule this content” and wait a few minutes. Claude reads the draft, uploads images to the right platforms, schedules social posts via Typefully, creates a WordPress post with the right categories and featured image, and sets up the email campaign in ActiveCampaign. All from one sentence.
Once, Claude messed up and sent an email that was blank. I told it to check my other scheduled emails and fix the system. It found a few other blank emails, loaded them properly and updated its rules so that after any scheduling task it logs in again and double checks nothing is blank.
As a human, I’ve made email send mistakes dozens of times, but was never able to fix it for good with a single sentence before.
Step 6: The feedback loop that makes everything compound
This is the part most people miss, and it’s the reason AI content systems get better instead of plateauing.
Every time you review a draft and change something, that change should become a permanent rule. Not just a one-time edit.
If you rewrite a paragraph because the AI used a word you’d never use, add that word to your “never use” list. If you restructure an opening because the AI buried the hook, add a rule about leading with the strongest sentence. If someone replies to your newsletter saying “I loved the part about X,” note that the specific, personal anecdote landed better than the general observation.
Over the first month, you’ll make a lot of corrections. By month three, you’re barely editing. By month six, the system produces content that sounds more consistently like you than you do on a bad writing day.
This is the compounding effect. Each correction is permanent. Each piece of feedback makes the next output better. A freelancer forgets your notes between assignments. An AI system with persistent context never does.
What’s more is any time somebody replies to one of my emails or comments on socials, I paste it into Claude and tell it to add it to the feedback doc. This has provided me dozens of improvements and future content ideas.
The above is an example of the feedback system I built in VSCode. I add the feedback in the Claude chat on the right, and the recent-feedback document (visible in the file manager on the left, and open in the center – with text hidden) gets updated in real-time.
Note the instructions: “Dump any comments, reactions, feedback, or notes here between session. These will be reviewed at the start of the next content session and used as light guidance (not rigid rules) when writing new content.” <– This came from me telling Claude I wanted a system to dump feedback or criticism my content received so I could improve in future, or come up with new content ideas.
The time math
Here’s what this actually costs in time:
Building the system (weeks 1-3): 15-25 hours total. Voice documentation, pillar definition, template creation, tool setup, distribution infrastructure. This is front-loaded and you only do it once.
Running the system (ongoing): 1-2 hours per week. Provide raw input (bullet points or voice memo), review drafts, approve scheduling, add feedback. Some weeks less if you batch.
Maintaining and improving: 1-2 hours per month. Update the voice doc, refine templates, adjust strategy based on what’s performing.
Compare that to either writing everything yourself (10-20 hours/week) or managing a freelancer or agency ($10,000/month plus your review time, which is often the same 3-5 hours anyway).
The system approach is cheaper and produces better output than both alternatives. But it does require the upfront build, and it requires someone technical enough to set up the AI tooling and distribution infrastructure. That’s the real barrier for most people.
For agencies: this is a service you can sell
If you run a marketing agency, everything I just described is a service offering waiting to happen.
Your clients already ask you for “content help.” Most agencies respond with blog writing packages or social media management. Those are commoditized and getting more so as AI makes basic content cheaper and faster.
The world needs less AI slop though, and most people know it. They just don’t want to pay humans to write content anymore, and they can’t be bothered to figure out how to build the system I’m describing.
We call that opportunity.
A full content system, the kind I just walked through, is a different category. It’s strategic, it’s personalized, it compounds over time, and it creates genuine lock-in because the system gets better the longer the client uses it.
You could build this for each client using the same framework, or you could white-label my system.
Price it at $2,000-5,000/month depending on volume and channels.
The AI does the heavy lifting on production. Your team handles strategy, quality control, and client relationships. That’s a high-margin service built on infrastructure that scales.
Or skip the build
Everything I just described is exactly what we built for Onfolio and what we now offer as a service called Parlance.
We do the voice documentation, the pillar strategy, the template creation, the distribution setup, and the ongoing production. You provide a few hours of input per month. We handle everything else.
If you read this guide and thought “this is exactly what I need but I don’t have 25 hours to build it,” that’s who Parlance is for.
If you read it and thought “I could build this for my clients,” we offer a white-label version for agencies.
Details and waitlist at onfolio.com/parlance
If this was useful to you, it would probably be useful to someone you know. Feel free to share it or forward it to a friend. One of the hardest things about being a small public company is simply being discovered, and word of mouth goes further than anything else.
This post reflects my personal views and experiences. It is not investment advice. For the most current information about Onfolio Holdings, refer to our SEC filings at sec.gov.