Joey Glyshaw is the founder of AlgoFuse, where he builds the shared AI layer that lets Claude and ChatGPT work from a company’s own data and procedures. He has sold on Amazon since 2015 and runs ZonFlip, an Amazon and TikTok Shop agency in Michigan, along with a small 3PL for Amazon sellers whose warehouse software he built himself with Claude.
A customer emails to ask why your product page says the tumbler holds 32 ounces when the box says 24. Nobody typed that number; it came from an AI draft that never saw your catalog.
An AI content strategy prevents that by settling what AI drafts, which files and data it reads on every run, and who signs off before anything ships. I run the PPC side of my agency from shared skill files that everyone’s Claude reads through a single MCP (Model Context Protocol) server, and content belongs in that same setup.
Here is how to build it.
TL;DR
- An AI content strategy decides what AI drafts, what it reads, and who approves.
- Store brand voice and each content SOP as files every teammate’s AI loads.
- Connect real catalog, review, and ad data so drafts start from facts.
- Google accepts AI-assisted pages that add value; mass-produced filler breaks its spam policy.
- When an editor fixes the same mistake twice, fix the file instead.
Drafts drifting off-brand or off-spec?
Spend thirty minutes with me sorting which content jobs belong in one shared system. You keep the plan, client or not.
Why AI Drafts Go Off-Brand and Off-Fact
Most bad AI content traces back to two gaps: the model has no shared voice to copy and no access to your facts.
Picture the usual setup. Each person keeps a saved prompt, a pasted voice doc, and a few old examples in their own chat history, and every one of those copies drifts a little each month. Nothing connects the model to the catalog, so it fills missing specs with numbers that sound plausible.
The symptoms show up fast:
- Product descriptions that use three different names for the same feature.
- Specs, sizes, or ingredients that match no record you hold.
- Blog posts that make claims your support team has never heard of.
- Emails that switch between “you” and “our customers” depending on who drafted them.
- A new hire who has to rebuild everyone’s prompts from scratch in week one.
In a Short I posted in April, I said it plainly: “My team re-explained our entire business every single day.” A better model will not fix that, because the model was never the weak part. Nothing was sitting underneath it.
How an AI Content Strategy Differs From a Content Plan
An AI content strategy keeps everything a content plan has, then adds the files, data, and checkpoints that let a model draft safely.
Most guides on the topic describe it as using AI to plan, create, and optimize content. That describes the tools. Whether the output sounds like you and gets the facts right depends on what sits underneath: the files and data every draft starts from, and the gate every draft passes through.
Goals, audience research and an editorial calendar still matter as much as they ever did. What AI changes is who types the first draft and how many pieces a small team can get through review in a week.
Which Content Jobs Go to AI and Which Stay Human
Give AI the jobs where a weak first draft is cheap to fix, and keep people on the calls that cost trust or money when they go wrong.
Some content jobs need no AI at all. Publishing approved posts on a schedule, adding UTM tags, posting the approved social copy, and pinging the reviewer in Slack are fixed rules, so I run them as plain automations and keep the model out of them.
The same sorting applies across the business, and there is a fuller map of which tasks AI should touch if you want to go past content.
The Five Files a Content System Runs On
Every content system I would trust runs on a handful of plain text files, each with one job.
content/
voice.md
positioning.md
audience.md
skills/
product-description.md
blog-post.md
weekly-email.md
review-checklist.md- Voice file. How you sound, shown through 10 to 15 sentences taken from your best pieces, the words you use, the words you never use, and two before-and-after rewrites.
- Positioning file. What you sell, who it is for, the claims you can prove, and the claims nobody may make (no “clinically tested” unless you hold the test).
- Audience file. The questions customers ask in their own words, pulled from support tickets, reviews, and sales calls.
- Skill files, one per content type. The workflow for a product description, a blog post, or the weekly email: the goal, the structure, the rules, and two approved examples.
- Review checklist. What a person checks before anything ships: facts against sources, claims against the positioning file, metadata and links.
My rule for every skill file fits in one line: “Write down the goal, the format, the rules, the examples.” Once those four things are on the page, the AI stops guessing, and your team stops starting each morning from zero.
Anthropic ships its own take on the idea. In Claude’s Agent Skills, a skill packages one workflow as a folder Claude opens on demand, holding instructions plus any scripts or resources it needs, and organization owners on Team and Enterprise plans can roll a skill out to every user. On AlgoFuse, the skill files sit on your own server behind your MCP server instead, so the files stay yours whichever model reads them.
Teams that already keep prompt playbooks for their LLMs have most of the raw material. The difference is that a skill file gets loaded on every run, whereas a playbook gets copied in only when someone remembers.
Where the AI Gets Its Facts
A voice file makes drafts sound right; data makes them correct.
On every build, my first question is which system holds the data, and content is no different. A model with no access to your catalog will invent a capacity or a battery life, since language models predict likely words from training data instead of looking anything up. If that failure mode is new to you, here is a longer piece on why AI makes things up.
MCP is an open protocol that gives Claude, ChatGPT, and other AI apps a way to reach outside data and tools through one kind of connection (MCP documentation). For content, these are the sources worth connecting, roughly in order of how much guessing they remove:
Since February 2, 2026, Amazon Ads has run an open beta of an MCP server for Ads partners holding active API credentials, which lets an AI app pull reporting data with no custom build. On the builds we do, the other sources run through an MCP server on your own backend, and your team adds it to Claude once as a custom connector. I covered why teams are moving to MCP in more depth elsewhere; for content, the point is narrower.
When I wire a data source, it lands in one central database first, and the AI reads from there. Every draft then pulls from the same copy of the facts, and the reviewer checks one place when a spec looks wrong.
Want to watch a content system get built in the open?
Skool hosts the paid AlgoFuse community, where operators and I build voice files, skill files, and data connections together.
What Google and the FTC Allow in AI Content
The rules are shorter than most of the internet suggests, and they reward the same structure. None of what follows is legal advice; check anything specific to your category with counsel.
Google’s own guidance on generative AI content, updated October 1, 2026, and its page on AI features and your website cover most of what a content team needs. The FTC rule on reviews sits in 16 CFR Part 465, published in August 2024.
Read together, a working review gate covers most of what Google is asking for. The FTC line is the one with no gray area for an ecommerce brand: every review and testimonial comes from a real customer.
Set Up an AI Content Strategy in Eight Steps
Build it in this order, since each step leaves you with a file or a number you keep.
Step 1: Name the Job Each Content Type Does
Write one line per content type: what it is for and the single number that shows it worked. Product pages convert, how-to posts earn search traffic, the weekly email drives repeat orders. A content type with no job gets cut before you automate it.
Step 2: Log Two Weeks of Time per Piece
Have everyone record time per piece by stage: research, drafting, editing, and approval. You need that baseline to know whether the system saved anything, and the log usually shows the hours sitting somewhere nobody expected.
Step 3: Write the Voice File From Your Best Work
Pull the ten pieces customers responded to most and mark the sentences that sound most like you. Turn those into examples, a word list, and a do-not-say list. A voice guide written from memory produces generic drafts, because the examples are what carry the voice.
Step 4: Turn Each Content Type Into a Skill File
Start with the type you produce most often. Write the goal, the structure, the rules, and two approved examples, then run ten drafts through it before you write the next skill file.
Step 5: Wire the Catalog First Through an MCP Server
Begin with the one source that removes the most guessing, usually the product catalog. Add reviews and support tickets once the first connection is stable, and give the AI read access only.
Step 6: Give One Person Sign-Off per Content Type
That person works from the review checklist and owns the call to publish. The checklist covers metadata too, since titles, meta descriptions, structured data, and image alt text show up in search results with your name on them.
Step 7: Pilot One Content Type First
Run one content type through the system for a month next to the old way. Compare hours per piece, edits per draft, and the number from Step 1 before you add a second type.
Step 8: Fix the File When You Fix a Draft
When an editor corrects the same thing twice, the correction goes into the voice file or the skill file, and every future draft inherits it. Name one person as owner of the files so changes land there instead of in a chat thread. Teams that only ever edit drafts end up typing faster with the same quality problems.
How to Judge an AI Content Tool
Tool lists age fast, so I judge any content tool, including the ones I use, by whether it fits the structure above.
- Files. Does it load your voice and skill files on every run, or does each user paste them in?
- Sharing. Can the whole team work from one set of files, with one person controlling edits?
- Exit. Can you export the files, prompts, and drafts in plain text if you leave?
- Data. Can it read your own catalog, reviews, and tickets through MCP or an API, or only the open web?
- History. Does it record who generated, edited, and approved each piece?
- Pricing. Is it priced per seat, per word, or per article, and what does the bill look like at three times your current volume?
A tool that stores your brand voice inside its own settings page is renting your voice back to you. Keep the files where you can move them, and treat the tool as one of several things that can read them.
Which Numbers Show the System Is Working
Track time and quality first, since traffic moves slower and depends on far more than drafting speed.
Do the arithmetic with your team’s figures. Say three people each give four hours a week to product copy and the system brings each of them down to two: that frees six hours a week, or 312 hours a year. Multiply those hours by what an hour of that work costs you fully loaded, because the result depends entirely on your hours and your rate.
Who Should Hold Off on an AI Content Strategy
The catch is that the system costs time before it saves any.
- Writing the voice, positioning, and first skill file takes real hours from someone who knows the brand well.
- Review time does not disappear; it moves from rewriting drafts to checking them.
- You pay Anthropic or OpenAI directly for every Claude or ChatGPT seat that drafts.
Wait a while if one of these fits:
- You publish fewer than four pieces a month, and one good writer with a saved prompt keeps up.
- Nobody on the team will own the review gate.
- You sell in a category with regulated claims, such as supplements, and have no expert to check drafts.
- The goal is volume for its own sake, which is exactly what Google’s scaled content policy targets.
Frequently Asked Questions
Quick answers to the questions owners bring up about running content through AI.
Keep the Voice File When the Models Change
Content that sounds like you comes from files that describe you, data that belongs to you, and a person who signs off. New models arrive every few months, and the voice file, the skill files, and the data connections are the parts that carry over.
Tools change. Models change. The structure stays.
Keep those files on your own drive or server, and every writer, freelancer, and model you bring in later starts from the same place. Building the data side with your team is what our AI integration engagements cover.
Ready to run your content from one shared system?
Bring your current voice doc and one content type to a call of thirty minutes, and you leave with a plan for it either way.
Leave a Reply