Same automation runs every morning. AI-generated articles → WordPress → Buffer → Pinterest. Roughly 200 scenarios fire per week between MikiMoneyAI (where you are now) and Solvara.
It took me eight months to build prompts that don’t break.
If you build automation in Make.com, this post will save you the eight months I just spent learning the hard way.
Part 1: What “AI Content Factory” Actually Means
Before I get into the prompts, here’s the architecture so the rest of this makes sense.
A single article on MikiMoneyAI is built by a Make.com scenario that does roughly this:
- A scheduler module fires at a set time (multiple times per day)
- An OpenAI module pulls a topic from a content calendar in a Google Sheet
- A Perplexity module pulls real-time data and recent news on the topic — this is the anti-hallucination layer
- A Claude module writes the article using the Perplexity grounding + my brand voice prompt
- An OpenAI module generates SEO meta data — title, meta description, slug, OG tags
- A Pexels module pulls a relevant featured image
- A WordPress module publishes the article as a draft, then promotes to live
- A Buffer module schedules X and LinkedIn posts pulling from the article
- A Pinterest module generates pin descriptions and posts the visual
The whole thing takes about 90 seconds end-to-end and costs me roughly $0.04 in API calls per article.
The architecture isn’t the hard part. The hard part is the prompts.
Part 2: Why Generic AI Prompts Don’t Work in Automation
If you’ve tried to plug a “100 ChatGPT prompts” pack into a Make.com scenario, you already know what I’m about to say.
Generic prompts return prose. Beautiful, well-structured, conversational prose. With markdown headings. And asterisks around words for emphasis. And bullet points formatted with em-dashes.
That works great when you’re chatting with ChatGPT in a browser tab.
It breaks immediately when the next module in your scenario expects clean structured data.
Here’s what actually happens:
- Your WordPress module gets confused by
**bold**markdown and publishes literal asterisks - Your Buffer module hits the X character limit because the AI ignored your instruction
- Your Pinterest description has emoji that break the API call
- Your meta description has line breaks that break the schema markup
- Your iterator can’t split the response because the AI ignored the format you asked for
I lost weeks to this before I figured out what was actually wrong: prompts written for chat are fundamentally different from prompts written for automation.
Part 3: The Delimiter System (My Single Biggest Unlock)
The fix is a three-character pattern that will save you months of pain.
When you ask Claude or GPT to return multi-section output, instruct the model to put |||SPLIT||| between every section you want to parse downstream.
Here’s the actual instruction I add to almost every production prompt:
Output your response with |||SPLIT||| between each labeled section.
Do not use markdown headings, asterisks, italics, or any text
formatting. Plain text only between delimiters.
Format:
Title |||SPLIT||| Meta Description |||SPLIT||| Body |||SPLIT|||
Tags |||SPLIT||| Featured Image Alt Text
The output looks like this:
The 5 ETFs Beating the S&P in 2026 |||SPLIT||| Discover the top 5
ETFs outperforming the S&P 500 this year, including dividend
strategies and sector tilts that worked. |||SPLIT||| [the full
article body here] |||SPLIT||| etf, investing, dividend stocks,
finance, market trends |||SPLIT||| A line chart showing five
ETF tickers outperforming the S&P 500 index in 2026
Now in my next Make.com module, I use a Text Parser or Iterator to split on |||SPLIT||| and each section maps cleanly to its destination field. Title goes to WordPress title. Meta description goes to AIOSEO. Body goes to post content. Tags go to WordPress tags. Alt text goes to the featured image.
No regex. No string manipulation. No 2 AM debugging sessions trying to figure out why the iterator returned an empty array.
I have variants for different downstream use cases:
|||IMGSPLIT|||— between image URLs and their alt text|||TAGSPLIT|||— between individual tags (parses straight into WordPress tags)|||METASPLIT|||— between SEO title, meta description, slug, and OG description
Same logic, different downstream parsers. One small pattern. Everything downstream gets simpler.
This single technique fixed about 80% of the broken scenarios I had when I started.
Part 4: The Anti-Hallucination Layer
The second biggest problem with generative AI in automation is hallucination.
The model confidently invents Amazon product ASINs that don’t exist. It cites studies that were never published. It quotes experts who never said the things attributed to them. In a chat window this is mildly annoying. In automation, it ships straight to your live blog and your audience reads it.
I lost trust with early MikiMoneyAI readers because of this. Articles went live citing fabricated statistics. Affiliate links pointed to dead Amazon products. It was bad.
The fix is what I call the Perplexity-grounded prompt pattern. The flow:
- Make.com calls Perplexity AI’s Sonar model with a query about the topic
- Perplexity returns current real-world data with sources
- That data gets passed as context into the Claude or OpenAI module
- The generation prompt explicitly says “use only the facts from the grounding data above; if you don’t know something from the grounding, say so or omit it”
The exact instruction:
Below is real-time research data with sources. You will write the
article using ONLY the facts present in this grounding data. Do not
invent statistics, quotes, product names, prices, or sources. If a
fact you want to mention is not in the grounding data, omit it
entirely. Do not hedge with phrases like "studies suggest" — only
reference research that appears in the grounding above.
GROUNDING DATA:
{{Perplexity output here}}
WRITE THE ARTICLE:
{{topic and brief here}}
This single change cut my hallucination rate from “almost every article had at least one fabricated stat” to “I haven’t caught one in six weeks.”
The Perplexity API is also dirt cheap — about $0.005 per query. Adding it to the pipeline costs me maybe $1.50/week across both brands. Worth every penny.
Part 5: The Errors That Almost Made Me Quit
If you build in Make.com long enough, you will hit these. Save yourself the time.
1. The IML Parser Error. Make.com uses its own templating language for variable interpolation. If your AI output contains characters Make’s parser interprets (like {, }, certain quote styles, or unescaped backticks), the next module crashes with a cryptic IML error. The fix: explicitly instruct the model in the prompt to never output curly braces, backticks, or smart quotes in its response.
2. Buffer Rate Limit + Retry = Duplicate Posts. Buffer rate-limits aggressively. When your scenario hits the limit and Make.com retries, you can end up with the same article posted to WordPress twice because the WordPress module ran successfully but the scenario looped back from a downstream failure. The fix: put the WordPress module after the social posting modules, not before. If Buffer fails, the post never publishes. Cleaner failure than duplicate posts.
3. Pinterest URL Validation. Pinterest is strict about destination URLs. If the AI generates a description with a UTM-tagged URL containing certain characters, the API returns a 400 error and the entire scenario stops. The fix: never let the AI generate URLs. Pass URLs as static variables from your sheet, and only let the AI generate descriptions and alt text.
4. Beehiiv 401 Authentication. Beehiiv’s API returns a 401 if your API key has any whitespace around it (super easy to introduce when copying from their dashboard). Took me four hours to figure that out. The fix: trim whitespace on every API key field, and store keys in Make’s connection settings rather than as text in modules.
5. WordPress Schema Validation. AIOSEO and other SEO plugins validate schema markup, and if your AI-generated meta description has a line break in it, the schema breaks and your post can’t be saved. The fix: in your meta description prompt, explicitly say “single line, no line breaks, no carriage returns.”
These five errors alone account for about 80% of the support time I spent in my first three months on Make.com.
Part 6: What I Wish I’d Had on Day One
If someone had handed me a document on day one that included:
- The delimiter system
- The Perplexity-grounded prompt
- A library of production-tested prompts for content, social, email, SEO, and data tasks
- Module recommendations for each prompt
- A debugging cheat sheet for the most common Make.com errors
…I would have saved roughly six months of building, breaking, and rebuilding.
So I wrote it.
The Make.com AI Prompt Vault
I just published The Make.com AI Prompt Vault on Gumroad. It’s a 46-page PDF containing:
- 50 production-tested prompts across 7 categories: Content Generation, Social Media, SEO/WordPress, Email/Newsletter, Data Processing, Customer Operations, Multi-Agent Workflows
- The Delimiter System chapter — the full pattern, variants, and how to wire it into Make.com modules
- Module recommendations on every single prompt (Claude API, OpenAI, Perplexity, HTTP, Iterators)
- The Top 10 Make.com Errors cheat sheet — what causes each, what fixes each
- The Perplexity-grounded prompt — the full anti-hallucination pattern, ready to paste in
- A 1-page Quick Reference companion — module recommendations, delimiter syntax, error codes, all on one printable page
It’s the document I wish I’d had on day one. Every prompt has been broken, rebuilt, and battle-tested running real scenarios across MikiMoneyAI and Solvara.
$39, instant download, lifetime updates, 30-day no-questions refund.
→ Get The Make.com AI Prompt Vault
A Final Note on Building This
I’m not a developer. My background is in utility billing operations — about as far from AI automation as you can get. I learned Make.com the same way most people will: by Googling errors at 11 PM, getting frustrated, fixing one thing, breaking two more, and slowly building intuition.
The reason I’m releasing the Vault as a paid product instead of a free PDF is because the prompts in it represent genuine compounded learning. Each one is the version that finally worked after the version before it broke. That’s worth $39 to someone who wants to skip the eight months I just lived through.
If you’re building anything with AI in Make.com — content, social, email, data processing, customer ops — and you’re still in the “everything is on fire” phase, the Vault is for you.
If you’ve already built a clean automation factory and you’re just here to read about how someone else does it, I hope this post helped, and I’d genuinely love to hear what your stack looks like in the comments below.
Either way: the future is automated, the tools are accessible, and the eight months I spent figuring this out are now compressed into 46 pages.
Build something.
— Mika
Editorial note: MikiMoneyAI publishes a mix of original and AI-assisted content, all reviewed and edited for accuracy. Long-form posts about the automation systems themselves — like this one — are written by Mika directly. We disclose AI assistance because trust matters more than performance.
Want more posts like this? Subscribe to the MikiMoneyAI newsletter for weekly breakdowns of AI tools, automation playbooks, and the income systems I’m building in real time.
