What AI Writing Instruments Get Flawed (And The Stack I Use As an alternative)

AI writing instruments make the writing half quicker, however writing was by no means the arduous half.

The arduous half in content material advertising and marketing is the knowledge—concepts, verified information, and reference materials. And that’s precisely the place these instruments fall quick.

I realized this after producing 40 articles by Claude. I’d tried the writing instruments first, however they only couldn’t deal with the half that really issues. And by “AI writing instruments” I imply the platforms constructed on prime of LLMs—Jasper, Frase, Writesonic, that class. What I used as an alternative was the LLM immediately, with my very own information and course of round it.

On this article, I’m sharing the 5 issues I bumped into and the way I deal with them now.

I’m not naming the precise instruments I examined. They’re not unhealthy merchandise. In case you don’t have robust writing or search engine optimisation abilities, otherwise you don’t have time for a extra hands-on course of, they’re a wonderful alternative. That content material is best than no content material. However if in case you have the abilities and wish to push high quality, they turn into the ceiling, not the flooring.

1. The analysis drawback: AI “analysis” simply recycles no matter’s already rating 

Most AI writing instruments “fact-check” the content material they generate by cross-referencing it in opposition to no matter ranks on Google.  Competitor advertising and marketing pages. Outdated weblog posts. Articles that copied their information from different articles. In observe, they’re laundering errors by consensus—if three fallacious sources agree, the AI treats it as truth.

And that’s a straight path to worldwide meta-spam.

I imply, after I let writing instruments deal with analysis, I obtained fallacious costs, incorrect options, and database numbers off by tens of millions. More often than not, it simply pulled from biased sources and had no solution to know they had been unhealthy. 

One of many instruments I examined used Gemini Deep Analysis because the article foundation. However Gemini—and I believe each different AI assistant—does the identical factor.

Once I wrote a comparability overlaying eight merchandise, I wanted eight separate fact-checked paperwork, one per product, plus a mode information, an modifying guidelines, and a immediate with required components. That’s 15-20 information I wanted the AI to reference all through the method. No writing software I examined may deal with that.

My resolution: all the time construct your individual reference information 

Construct verified information information for each product and competitor you cowl. Begin with a information base in your personal merchandise, in a type the place you possibly can simply generate paperwork from it: pricing, options, use instances, all the important thing numbers. I truly vibecoded a software for that.

If it’s worthwhile to characteristic rivals in your content material, put together paperwork for the components you need referenced: their pricing pages, characteristic lists, limitations, and so on. I downloaded competitor touchdown pages, took screenshots, and vibe-coded a scraper to tug pricing and options from official sources.

By no means begin any AI content material mission till your information information are accomplished. In case your mission is supposed to take 4 weeks, use three weeks for these information.

2. The method drawback: writing instruments attempt to one-shot your article, however good writing doesn’t work that method 

Writing instruments are meeting traces: configure inputs, press generate, gather output. However writing is nearer to cooking—you style at each stage, add some unplanned components, or possibly flip the factor into one thing else. 

It doesn’t matter how a writing software handles model voice. Whether or not it’s a dropdown, a mode file, or a set of directions, the end result all the time wants modifying. Getting our voice proper took 5 – 6 rounds per article. I’d learn a draft again and say “that appears like a press launch” or “put the quantity first, you’re burying the lead.” You want a dialog for that.

That is additionally an interface drawback. Modifying AI-generated textual content means working at each degree: rewriting a single sentence, restructuring an entire part, fixing a sample throughout all the article. In a chatbot, I simply requested for what I wished in plain English. Writing instruments gave me mounted modifying choices that couldn’t deal with that vary.

My resolution: break your course of into repeatable prompts or abilities 

Break your workflow into repeatable duties and develop prompts for every:

• Reality-checking.
• Inside consistency checking.
• Model and construction enforcement.
• Product positioning enforcement.

Trial and error till every immediate nails it.

Afterward, these prompts can turn into your Claude skills, if/while you determine to make use of automated content material workflows.

Tip: For an important steps, I ran my prompts twice, or ran the identical examine by a second AI to catch something the primary one missed.

3. The dimensions drawback: writing instruments deal with each article as an island 

Writing instruments encourage you to consider automating content material at scale. Some even supply workflow options for it. However I discovered them irritating in observe: arduous to construct, human-in-the-loop performance may be very restricted, and the output drifts the extra nuanced your necessities get. 

AI assistant already solved this, and Claude Code took it to the following degree. I may kind “scan each article for Product X’s pricing and examine it in opposition to the reference file” and it will do it. When one thing wanted adjusting, I simply instructed it.

That’s performance that writing instruments don’t supply, though the underlying LLM is able to it.

My resolution: get used to working with Claude Code 

In Claude Code and OpenAI Codex, one instruction kicks off the entire course of. Tt fetches search engine optimisation information, pulls from my reference information, grabs what it wants from the net, and writes the article in phases. I outlined the phases, then let it run whereas I did one thing else. 

That is additionally the place analysis instruments plug in. MCP integrations like Ahrefs’ let you pipe real data directly into these workflows—we’re experimenting with a full Claude Code pipeline where SEO research happens automatically. If your tool doesn’t support MCP yet, pull the data manually. Even screenshots work, as long as you give the AI specific data to work on.

4. The economics drawback: when the wrapper prices greater than the engine 

A chatbot subscription prices $20 a month and offers you the most recent mannequin with no article or phrase limits. The writing instruments I examined value $50-200 a month, one even $2k a month, and ran older fashions with caps on how a lot you can generate. Seems like paying extra for much less.

Right here’s an instance. To jot down one of many articles for the experiment, I pulled the top-cited articles for my key phrase (utilizing Ahrefs’ Model Radar), then had Claude undergo these pages to extract the construction and use that as a top level view template for content material technology. Then I requested it to weave in my very own concepts. Analysis, construction, writing—multi functional dialog, controlling each stage.

However possibly I’m fallacious. Perhaps a writing software with all the things on board is extra your model. I’ll go away it to you to determine what makes extra sense economically. I don’t wish to inform you what to do together with your cash, however I do know that for my wants, I’m by no means going again to AI writing instruments.

There’s additionally one thing a bit self-defeating concerning the AI software ecosystem. Each time an LLM supplier releases a greater mannequin, most of the instruments constructed on prime of it lose a part of their motive to exist.

My resolution: make investments extra in what you feed the AI

Redirect money and time towards:

• Analysis instruments that go deep. Wealthy key phrase information, search intent analysis, competitive gaps, AI-preferred content formats, etc. Writing tools bolt on a surface-level version of this. Dedicated platforms have years of infrastructure behind them (here’s ours).
• Your editorial system. Prompt libraries, fact-checking workflows, style enforcement, Claude or Codex skills. The stuff that keeps your judgment in the loop at every stage. Same principle as the reference files: invest in the inputs.

5. The content material technique drawback: writing instruments provide you with one course of for 2 very completely different jobs 

Writing instruments assume all content material works the identical method. Feed it a key phrase, get an article. However I see content material splitting into two tracks in our line of labor, and writing instruments can’t deal with both one correctly.

The primary is searchable content material. Product documentation, assist articles, comparability pages—the stuff most groups handled as a chore. It’s all of a sudden vital as a result of if an AI mannequin can’t floor its reply in one thing you printed, it’ll use no matter it finds. Or hallucinate. Your product documentation is your model’s voice inside each AI dialog now.

Right here’s what that appears like when it really works. I requested AI Mode, “What number of manufacturers are you able to observe in Model Radar?”, and it cited our docs immediately.

And right here’s what occurs when there’s a spot: no official supply cited. Fortunately, the truth that I requested AI mode about obtained talked about in one other piece, however that was nearly accidentally.

The second, I feel, is shareable content material. Actually human-first content material. Stuff that comes from private expertise and may’t be templated. My AI misinformation experiment is an instance: it ranked for nothing, however drove 24k visits and extra social traction than I may rely.

My resolution: select flexibility over comfort

Each content material tracks want completely different approaches, and AI chatbots are the one instruments versatile sufficient to deal with each. So what you want is a course of for creating documentation that you could simply share with AI.

For searchable content material, audit your product documentation and assist content material. If an AI mannequin can’t reply a primary query about your product utilizing your individual content material, that’s a spot another person will fill, unintentionally or intentionally.

You’ll be able to chat with the preferred AI assistants to identify holes, or arrange monitoring in a software like Ahrefs Brand Radar to do it at scale.

For shareable content, build an idea pipeline. Start a scrapbook. Store ideas, facts, quotes, social posts, newsletter excerpts, and anything you might want to give AI access to later.

You can use Notion, Evernote, whatever suits you. But consider vibecoding a custom tool, like my colleague Louise. That method, you possibly can bake in options like an “instance finder” that surfaces related assist for claims in your writing, or simply generates content material concepts out of your materials on the spot.

One other thought: arrange an AI agent that scours the net for content material concepts on a schedule. I constructed one with Relay that goes by LinkedIn and Reddit conversations (truthful use) each 7 days. It helped me keep on prime of all the brand new content material popping out quicker than ever and keep sane.

If you wish to hold a fixed pulse on new content material in your area, strive our new software, Firehose. It streams the net in actual time on any subject you outline, with superior filtering. You describe what you’re on the lookout for in pure language, and it’s able to go. You too can join it to your AI brokers by the API.

Closing ideas

Thanks for studying! When you have any questions or feedback, let me know on LinkedIn.