How to Build a Ghostwriting AI Agent in 2026: Complete No-Code Tutorial (Make.com + ChatGPT)
I didn't build my first ghostwriting AI agent to get rich quick. I built it because I was drowning in drafts. After 18 months of testing, failing, and rebuilding, I finally have a no-code system that researches, outlines, and drafts human-sounding articles while I sleep — and I'm going to show you exactly how to build it.
I'm Aymane Jabel. I test AI workflows from a small office in Casablanca. In 2024, I tried to scale a blog by hiring writers. I attempted to manage them with Trello and Google Docs. I failed completely after three months — the voice was inconsistent, the fact-checking was weak, and I spent more time editing than writing.
Then I tried the opposite: I relied on one-click AI writers. I achieved speed, yes, but I also achieved something else: three AdSense policy warnings and zero reader trust. The content was generic, outdated, and clearly not written by a human who had actually tested anything.
So I learned to stop looking for a magic button. What I needed was an agent — a workflow that could do the boring research and structuring, but leave the final judgment, experience, and editing to me. I tested 11 different versions in Make.com. The first 7 failed. Version 8 succeeded for two weeks, then broke when OpenAI changed its API. Version 11 is the one I still use today.
This is the complete, step-by-step build.
what is a ghostwriting ai agent
A ghostwriting AI agent is not a chatbot. It's not a single prompt you paste into ChatGPT. It's a multi-step automation that acts like a junior researcher and ghostwriter working under your direction.
Think of it like this: you give it a keyword in a Google Sheet. The agent then wakes up, searches the live web for 2026 sources, extracts real facts and quotes, builds an outline based on search intent, writes a first draft in your voice, adds your personal testing notes, and saves it as a WordPress draft for you to review. It never publishes by itself.
The key difference from old AI tools is memory and tools. A classic LLM just predicts the next word. My agent uses tools: Perplexity for live research, Google Sheets for memory, and Claude for long-form writing. I learned this after I attempted to make ChatGPT do everything in one prompt. It failed because it hallucinated dates and invented studies. When I split the work — research first, then outline, then write — I finally succeeded.
I use this agent for my own blogs, not for clients. It does about 70% of the grunt work. I do the final 30%: adding my own photos, my failed tests, my opinions, and fact-checking. That's the part Google calls E-E-A-T, and it's non-negotiable in 2026.
why build no code ai agent 2026
You might ask: why not just use Surfer AI or Koala or any 2026 writer? I tested them all. I tried KoalaWriter in January, then NeuronWriter, then three newer tools. I achieved fast drafts, but I failed to achieve control. I couldn't force them to cite real Moroccan data, I couldn't inject my personal testing history, and I couldn't make them follow my exact outline structure.
Building your own in 2026 is cheaper and smarter for three reasons.
First, cost. Make.com costs $9/month. Perplexity API is $5 for 1,000 searches. OpenAI and Claude APIs cost me about $0.18 per 1,500-word draft. That's under $20/month for 30 articles if you review them yourself. I previously relied on a $99/month tool that limited me to 25 articles.
Second, customization. My agent knows I write from Morocco. I baked into the prompt: "Always mention if a tool is available in North Africa, and add a personal testing note." No SaaS tool lets you do that deeply.
Third, safety. In 2025, Google's helpful content updates penalized pure AI sites. I learned that the hard way. An agent that forces human review, adds real sources, and includes first-person experience is what I now rely on to stay AdSense safe. I don't want autopublish. I want autoprep.
When I finally succeeded in March, I didn't celebrate the speed. I celebrated the consistency. Every draft now has the same research depth, the same citation format, and the same place for my personal input.
tools needed free
You don't need to code. You need accounts. Here is my exact stack, and the free tiers work to start:
1. Make.com (free plan) – The workflow engine. This runs the scenario.
2. Google Sheets (free) – Your control panel. One row = one article idea.
3. Perplexity API ($5 credit) – For live web research with sources. This is critical. I tested ChatGPT browsing and it failed for citations. Perplexity returns clean JSON with URLs.
4. OpenAI API (pay-as-you-go) – I use gpt-4o-mini for outlining. Cheap and fast.
5. Anthropic Claude API – I use Claude 3.5 Sonnet for the actual ghostwriting. It sounds more human in my tests.
6. WordPress with Application Passwords (free) – To create drafts automatically.
Optional: A webhook trigger if you want to start it from your phone.
I achieved my best results when I stopped trying to use one LLM for everything. I now use Perplexity to research, OpenAI to structure, and Claude to write. That three-brain system is what fixed my hallucination problem.
how to build ghostwriting ai agent step by step
I'm going to build this exactly as I did. Follow in order. Don't skip step 1 — my first six failures came from a messy sheet.
step 1 create keyword sheet
Open Google Sheets. Create a new sheet called "AI Agent Control". I tried using Airtable first. I attempted Notion. I failed because Make.com connects fastest to Sheets.
Create these exact columns in row 1: keyword | search_intent | target_audience | status | personal_note | draft_url | sources
Here's how I use them: In 'keyword' I put "best portable solar panels 2026". In 'search_intent' I write "commercial investigation". In 'target_audience' I write "Moroccan van lifers". In 'status' I write "ready". The 'personal_note' is crucial for E-E-A-T. I write: "I tested the EcoFlow 110W in Merzouga desert, it failed in wind." That note gets injected into the final article.
I learned to never leave personal_note empty. When I relied on AI alone, the content felt fake. When I added one real sentence, my time-on-page doubled.
Add 5 test rows. Set only one row to "ready" for now. This is your queue.
step 2 setup make.com scenario
Log into Make.com. Create a new Scenario. Click the big plus, search for "Google Sheets - Watch Rows". This is your trigger.
Connect your Google account. Select your sheet, choose "Watch Rows". Set it to trigger when 'status' equals "ready". Set limit to 1. I tested running every 15 minutes.
This module will pull your keyword and personal_note. I failed here twice because I forgot to set the filter. The scenario ran on every row and burned my credits.
Next, add a "Tools - Set Variable" module. Name it 'article_context'. Combine your fields: "Keyword: {{keyword}}. Audience: {{target_audience}}. My personal test: {{personal_note}}. Write from Morocco perspective."
This context variable is what I succeeded with after 4 attempts. Previously I passed fields separately and the AI lost context.
Add a Router after this. Route 1 goes to research, Route 2 goes to writing. This separation is what I learned from prompt engineering experts in 2026: give each AI one job.
step 3 connect perplexity api
This is the most important upgrade I made in 2026. I used to rely on ChatGPT browsing. It was slow and made up sources.
In Make.com, add HTTP - Make a request module. Method: POST. URL: https://api.perplexity.ai/chat/completions
Headers: Authorization: Bearer YOUR_KEY, Content-Type: application/json
Body (raw JSON):
{"model":"sonar-small-online","messages":[{"role":"user","content":"Research '{{keyword}}' for 2026. Return 5 key facts, 3 pros, 3 cons, and 5 source URLs with titles. Audience: {{target_audience}}."}]}
I tested this prompt 20 times. When I asked for "article", it summarized. When I asked for facts and URLs in JSON, it succeeded every time. The output gives you real, clickable sources.
Add a JSON parser after. Map the response to a variable called 'research_data'. I store the sources in the 'sources' column of my sheet. This is my proof of research for E-E-A-T. I learned that Google doesn't punish AI content, it punishes content without evidence.
step 4 claude prompts
Now the ghostwriting. I split this into two prompts — outline first, then draft. I attempted one-shot writing and failed. The articles were messy.
Module 4A: OpenAI - Create Completion (for outline)
System: "You are an SEO strategist. Create a detailed H2/H3 outline for 2026." User: "Context: {{article_context}}. Research: {{research_data}}. Return ONLY a markdown outline with 5-7 H2s."
Module 4B: Anthropic Claude - Create a Message (for writing)
This is where the magic happens. My prompt after 30 tests:
You are Aymane, a tester from Morocco who writes in first-person. Write a helpful, conversational blog post using this outline: {{outline}}. You MUST integrate this personal experience naturally: {{personal_note}}. Use the research facts: {{research_data}}. Cite 3 sources inline like [Source]. Write in simple English. Add one "What I learned" paragraph near the end. Do NOT sound like AI. Do NOT promise results.
I achieved a human voice when I forced Claude to use my personal_note verbatim. I also learned to add "Do NOT sound like AI" — it reduces fluff.
Set max tokens to 4000. Temperature 0.7. I tested temperature 0.9 and it got too creative and failed fact-checking.
step 5 wordpress integration
Final step: don't publish, just draft. In Make.com, add WordPress - Create a Post.
Connect using Application Passwords (Users → Profile in WordPress). Map Title = {{keyword}}. Content = {{claude_output}} + "
Sources
- {{research_data.sources}}
Reviewed in Casablanca, May 2026. AI-assisted draft, human edited.
"Status: Draft. Category: choose yours.
Then add a final Google Sheets - Update a Row module. Set 'status' to "drafted" and 'draft_url' to the WordPress link. This stops it from running again.
I succeeded when I added that last update step. Before, the agent kept re-processing the same keyword and I burned $12 in one night.
Test it: set one row to "ready", click "Run once" in Make.com. Wait 60-90 seconds. Check WordPress → Drafts. You should see a full article with your personal note, real sources, and a disclaimer.
That's it. The whole system runs while you add keywords. I now spend my time testing products and adding one-sentence notes, not writing from scratch.
best prompts for ghostwriting ai
I tested over 100 prompts. These three work best in 2026:
1. Research Prompt (Perplexity): "Find 2026 data about [topic] for beginners in [country]. Return facts, not opinions. List 5 URLs from last 6 months." I learned to add "not opinions" — it reduced blogspam sources.
2. Voice Prompt (Claude): "Rewrite this in my voice: I'm Aymane from Morocco, I test things myself, I tried X and failed, I use simple words, short sentences." I achieved consistency when I gave it my failure history.
3. E-E-A-T Injector: "Add a personal testing paragraph after intro: 'I tested this for 7 days in Casablanca. Here's what worked and what didn't.' Use my note: {{personal_note}}." This is what I relied on to pass manual reviews.
The key I learned: never ask AI to "write a blog post about X". Always give it research, outline, voice, and a personal fact. That's real prompt engineering.
common mistakes
I made all of these. Please avoid them.
1. Auto-publishing. I attempted this in February. I failed. One article hallucinated a Moroccan law. I got an email from a reader. Now everything is draft-only.
2. No personal notes. When I relied only on AI, my bounce rate was 89%. With one real test sentence, it dropped to 54%.
3. One LLM for all. ChatGPT is bad at live research. Claude is bad at JSON. Perplexity is bad at long writing. Use the right tool.
4. Forgetting sources. I achieved AdSense approval only after I started automatically appending source lists. Google wants evidence.
5. Complicated Make scenarios. My first build had 19 modules. It broke constantly. My current one has 7 modules and has run for 63 days straight.
make ai content adsense safe
This is not legal advice, but this is what I do after two AdSense reviews in 2025-2026.
First, transparency. Every draft gets a footer: "AI-assisted research, human tested and edited in Morocco, May 2026." I learned that hiding AI use is riskier than disclosing it.
Second, add original value. My agent cannot go to Merzouga desert. I can. I add one photo, one test result, one local price in MAD. That's what Google calls Experience.
Third, cite everything. My Perplexity step forces 3-5 real URLs. I keep them in the article. I achieved a 100% originality score on Originality.ai after adding citations.
Fourth, edit for 10 minutes. I fix transitions, add my "I tried" and "I learned" phrases, remove fluff. Never publish raw AI output. I tested raw once. It ranked for 3 days then disappeared.
I don't make income claims. This is a productivity tool, not an ATM.
e-e-a-t checklist
Before I hit publish, I run this checklist — it's taped to my monitor:
- Experience: Did I add at least one personal test, photo, or failure? (I write "I tested in Casablanca")
- Expertise: Did I cite 3+ current sources from 2026?
- Authoritativeness: Is my author bio visible with real location?
- Trust: Did I disclose AI assistance and the review date?
- Helpfulness: Would this help a complete beginner take the next step?
I learned that E-E-A-T is not a plugin. It's a habit. My agent provides the structure, I provide the proof.
Building this ghostwriting AI agent took me four months of failing, but now it saves me 12 hours per week. I still write — I just write the important parts: my tests, my opinions, my lessons. The agent does the research heavy lifting. If you're a solo creator in 2026, you don't need a team. You need a system you control.
Start with step 1 today. Build the sheet. Don't overthink the perfect prompt — I tested dozens and still tweak them monthly. The real win is owning your workflow.
Written by Aymane Jabel
I test AI tools and side hustles with $0 budget and share real, unfiltered results. No hype, just honest experiments. More about me →
){kind=link}
0 Comments