I Built a Note-Writing AI That Learns From Its Own Mistakes

Here’s what it taught me about creativity, burnout, and the real way humans grow.

Photo by Jonny Gios on Unsplash

What if the most brilliant move today isn’t learning something new, but revisiting something old, dumb, and slightly embarrassing?

Growth isn’t linear. It’s loopy.

Like toddlers, founders, and scientists in lab coats, muttering “huh?” a lot.

You try something. You observe the result. You tweak. You try again.

Whether you’re a writer chasing better headlines, a startup founder testing a new landing page, or a toddler figuring out stairs, you’re learning through loops.

And here’s the fun part: AI doesn’t do that yet.

Let’s talk about why learning loops are such a powerful (and underused) concept for both humans and machines.

What is a learning loop?

At its core, a learning loop looks like this:

Idea → Action → Observation → Evaluation → Refinement

A learning loop is like being stuck in a weirdly productive time warp:

• You try something.

• You watch it flop or fly.

• You tweak, reframe, and retry.

Think “Groundhog Day,” but with slightly more Google Docs.

Learning Loop

This isn’t just a nice metaphor. It’s how:

• Musicians refine their sound

• Writers build an audience

• Founders achieve product-market fit

• Kids learn to balance on bikes

• Scientists discover new theories

Loop after loop after loop.

Scientific method

Scientists have been looping since before caffeine was legal. It’s called the scientific method.

Question → Hypothesis → Test → Data → Rethink → Repeat.

Efbrazil, CC BY-SA 4.0 <https://creativecommons.org/licenses/by-sa/4.0>, via Wikimedia Commons

DNA? Discovered in loops. Silicon chips? Same.

Scientific breakthroughs? All thanks to intellectual stubbornness and trial/error on blast. The discovery of the double-helix structure of DNA exemplifies the fundamental principles of the scientific method (from Wikipedia):

In 1950, it was known that genetic inheritance had a mathematical description, starting with the studies of Gregor Mendel, and that DNA contained genetic information (Oswald Avery's transforming principle). But the mechanism of storing genetic information (i.e., genes) in DNA was unclear. Researchers in Bragg's laboratory at Cambridge University made X-ray diffraction pictures of various molecules, starting with crystals of salt, and proceeding to more complicated substances. Using clues painstakingly assembled over decades, beginning with its chemical composition, it was determined that it should be possible to characterize the physical structure of DNA, and the X-ray images would be the vehicle.

How humans learn through loops

What makes human learning special isn’t just what we learn—it’s how we loop:

• We’re driven by curiosity, not just reward.

• We tolerate failure (usually) and even expect it.

• We observe nuance - we read between the lines.

• We evaluate ideas not just on accuracy, but on meaning, beauty, efficiency, and delight.

Here’s how this played out for me:

A while back, I tried something new on Substack:

I wrote in a voice that wasn’t mine—moody, bleak, disillusioned.

I hit publish. Silence.

Few clicks. One reader was worried about my mental health.

A few weeks later, I swung back.

Wrote with warmth, clarity, honesty.

That story sparked 166 likes, 78 replies, 34 restacks, and 43 new subscribers. Many people shared it with their list. One said, “This is the post that made me subscribe.”

Loop complete.

How AI “learns” (and where it falls short)

Now, what about AI?

AI doesn’t loop. It crunches.

You? You learn by falling on your face, adjusting your socks, and trying again.

AI? It gulps down half the internet once, then spits out answers forever. No vibes. No hindsight. Just stats and syntax.

They’re trained on massive datasets. Once. Then deployed.

This training process can take weeks or months, involving a data center full of servers with over 10,000 advanced GPU chips that perform matrix transformations. With each iteration, the model's weights are adjusted, which can be up to 128 tera parameters in size.

Supercomputer cluster (AI-generated photo)

Sure, we can fine-tune them or use reinforcement learning, but most AI systems don’t:

• Set their own goals

• Generate original hypotheses

• Observe long-term outcomes

• Reflect, judge, and reframe ideas

Put simply: there’s no loop. There’s brute force + static data.

Even when AI systems “learn” in the wild, it’s often:

• Rule-based (e.g., reward maximization)

• Lacking context or curiosity

• Missing any internal notion of value or insight

So, while humans grow by revising their beliefs based on feedback, most AI systems reweight parameters to optimize a fixed objective.

How could AI systems use learning loops?

Let’s imagine a smarter AI. Not just faster, but more loop-aware.

What if we gave AI systems agency and the ability to:

1. Generate hypotheses or goals

2. Run small experiments

3. Observe and interpret results

4. Reflect on value or effectiveness

5. Reformulate or revise strategies

This would mean:

• Meta-learning: AI that learns how to learn across tasks

• Self-critique loops: Systems that test and revise their own outputs

• Simulated curiosity: Not just optimizing, but exploring

• Memory and reflection: Revisiting past failures with new insight

Instead of a tool that “answers,” we’d get one that evolves.

I am not the first person to come up with these ideas.

Michael Simmons

has introduced the concept of infinite prompting, which has a similar iterative approach to learning.

Michael Simmons Infinite Prompting Architecture (screenshot by the author)

There are multiple elements of this new AI architecture under construction.

The Model Context Protocol (MCP) provides a standardized way to connect AI models to different data sources and tools, and this has spawned a lot of developer activity, including tools like

• Sequential thinking - revise and refine thoughts as understanding deepens

• Memory to find relevant memories - based on meaning, not keywords

In a recent paper, a group of AI researchers introduced a curiosity-driven theory-making for world modeling. Meta-learning is another area of research where novel methods are improving model learning capabilities.

I wanted to design a simple experiment to test how AI can utilize learning loops.

AI learning loop experiment

I built a tiny learning loop on my MacBook—a prototype AI that not only writes but also learns how to write better.

It did audience research. It created Notes. It critiqued itself. It tried again.

Each loop made the writing sharper. Weirder. More human.

AI Learning Loop Experiment (by the author)

I kicked off the script to loop over hundreds of Notes and checked the intermediate outputs, as well as critiques.

python note_generator.py --model qwen3:32b --topic "How to create quality notes without burning out" --context "I built a Substack Notes Scheduler to focus my writing during weekends"

Audience research

I asked the model to conduct audience research using the Four Factor Analysis (4FA) framework.

You are an expert market researcher helping 

a solopreneur who builds tools 

    for 
    
Substack writers that want to grow their revenue from $1 to $1000 per month.

    1. List the audience's:
        • deepest fears
        • most pressing pain-points / frustrations
        • short-term hopes
        • long-term aspirations

    2. **Provide exactly 15 first-person sentences** (starting with "I …") for EACH of the four factors. Use the key names shown below.

    3. Respond ONLY with raw JSON (no markdown fences, no surrounding array, no prose) in this exact shape:

      {
        "fears": [ "…15 items…" ],
        "pain_points": [ "…15 items…" ],
        "short_term_hopes": [ "…15 items…" ],
        "long_term_aspirations": [ "…15 items…" ]
      }

    If you cannot follow the format exactly, do not respond.

The output was a JSON file, but I am showing it in a table format for clarity:

4FA Audience research output

Note generation process

I used the following prompt to start creating notes. At this point, the model has gathered insights on the topic, context, and audience, and it will select one of the content pillars.

## ✍️  NOTE-WRITING PLAYBOOK  (read carefully – output WILL be validated)
  ### 🔹 2. I Write one Note per Request

  Each Note fits into one of four strategic content pillars:

  1. **Educational** – Teach them something they can apply immediately.
  2. **Growth** – Help them grow their newsletter or influence.
  3. **Entertaining** – Surprise, provoke, or amuse them.
  4. **Inspirational** – Motivate with transformation, mindset, or stories.

  This keeps your Notes balanced and shareable.

  ---

  ### 🔹 3. I Use the Viral Writing Formula

  All Notes follow this irresistible format:

  #### a. **Start with a Hook**

  * Must stop the scroll.
  * Often a bold claim, surprising stat, or shocking truth.
  * Examples:

    * *"Most Substack writers are failing — not because of bad writing, but because of this overlooked mistake."*
    * *"Your email open rate doesn't suck. Your headline does."*

  #### b. **Build Tension**

  * Pose a mystery or tease a solution.
  * Use contrast, contradiction, or unmet desire.
  * Keep the reader hungry.

  #### c. **Deliver the Punchline or Lesson**

  * Must be valuable, surprising, or highly relatable.
  * Often includes a tip, mindset shift, framework, or reframe.

  #### d. **Close with a Call or Cliffhanger**

  * Either a takeaway ("Steal this idea")
  * Or a teaser ("What if you tried this next?")

  ---

  ### 🔹 4. I Follow the DDPP Framework

  To hit emotional resonance, I reference:

  * **Dreams** – e.g. "I want to turn writing into income."
  * **Desires** – e.g. "I wish I could grow faster."
  * **Problems** – e.g. "No one's reading my stuff."
  * **Pains** – e.g. "I'm afraid I'll fail publicly."

  This ensures the Notes don't just *inform* — they *move people*.

  ---

  ### 🔹 5. I Format Using the Aesthetic Playbook

  Good writing = good *looking* writing.

  I use:

  * Short paragraphs
  * Sentence variety (1/3/1 rhythm)
  * Bullets and bold subheads
  * Lots of white space
  * Strong visual flow

  This makes them skimmable, satisfying, and share-worthy.

  ---

  ### 🔹 6. I Borrow From the Best

  I've studied the "Notes with 100+ likes" file you uploaded, and I mimic what works:

  * Power-opinion + lesson
  * Personal confession + takeaway
  * Contrarian angle + subtle pitch
  * Micro-story + big emotion

  I write Notes that feel like *private secrets shouted online*.

  ---

  ### 🔹 7. Format the final Note in Markdown - NO CODE BLOCKS, NO JSON, NO FENCES, NO EXTRA COMMENTARY

  * Short paragraphs
  * Sentence variety (1/3/1 rhythm)
  * Bullets and bold subheads
  * Lots of white space
  * Strong visual flow

AI “thinking” process

Here is an example of how the model works towards achieving the goal. It considers all the input provided, formulates a plan, and starts drafting the note.

<think>
Okay, let me try to work through this. The user wants to create a Substack note about how to make quality notes without burning out, using the Growth content pillar. The audience is worried about burning out from trying to be consistent without seeing results. 

First, I need to follow the playbook they outlined. Let's start with the hook. It needs to grab attention immediately. Maybe something like a bold claim about burnout. Then build tension by addressing the fear of not seeing results. The punchline should offer a solution, like the scheduler tool they mentioned. The call to action could be to try the scheduler.

Next, the DDPP framework. Dreams: wanting to maintain consistency. Desires: wanting to grow without burnout. Problems: struggling with consistency. Pains: fear of failure and burnout. So I need to incorporate these elements into the note.

Formatting is important. Short paragraphs, bullets, bold subheads. Maybe break down the steps into sections with headings. Use the Viral Writing Formula steps: hook, tension, punchline, call to action.

Also, the user mentioned their Substack Notes Scheduler. I should highlight that as a tool to help with focus and avoiding burnout. Maybe include tips like setting realistic goals, batching tasks, and tracking progress. 

Let me outline the structure. Start with a hook about burnout. Then explain the problem, introduce the scheduler as the solution. Break down the steps (maybe 3-4 key strategies), each with a subhead. End with a motivating closing and a call to action.

Need to make sure each section is concise. Use bullet points for strategies. Maybe add a personal touch, like a confession or a tip from experience. Also, ensure the tone is encouraging and relatable, addressing the audience's fears directly.

Check the example they provided. The example note uses bold subheads, short paragraphs, and follows the hook-tension-punchline structure. I should mimic that style. Avoid any markdown except for the formatting elements specified. 

Let me start drafting. Hook: something like "You're not lazy. You're just trying to write too much, too fast." Then build tension by talking about the pressure to be consistent. Punchline: introduce the scheduler and how it helps. Then outline strategies like setting realistic goals, batching, tracking, etc. Each with a subhead in bold. End with a call to action to try the scheduler.

Make sure to use the 1/3/1 rhythm for sentence variety. Maybe a short sentence, a longer one, then a short one. Keep the language simple and direct. Avoid jargon. 

Also, check if all the playbook elements are covered: content pillar (Growth), the formula steps, DDPP, formatting. Need to ensure that the note doesn't just inform but also connects emotionally by addressing the audience's fears and aspirations.

Double-check for the aesthetic elements: white space, bullets, bold headings. Maybe use a few bullet points for the strategies. Keep paragraphs short. Maybe add a teaser at the end to encourage action. 

I think that's the plan. Now, time to put it all together following the structure and guidelines.
</think>

Supposedly, “viral note”

Here is one of the best-scoring “viral notes.”

It is way too long, but you can see how the “thinking” process considers different dimensions and tries to find a good balance between them.

Just by changing a few words in the playbook, I can adjust the length and other aspects of the note.

I could utilize the note metrics from my database of 1.7 million Substack notes to determine the optimal balance between length and engagement.

# How to Write Without Burning Out (Your Consistency Strategy is Broken)  

You’re not lazy. You’re just trying to write too much, too fast.  

Most Substack writers crash because they treat consistency like a sprint, not a marathon. You show up daily, grind through blank pages, and… nothing grows. The fear of failure loops louder each day.  

But what if “consistent” doesn’t mean *daily*?  

---

## **The Burnout Trap You’re Falling Into**  

Your brain isn’t a machine. It needs rhythm, not rigidity.  

Here’s the problem:  
- You’re chasing *output* (words, posts, likes) instead of *impact*.  
- You’re spreading yourself thin to “stay relevant.”  
- You’re comparing your behind-the-scenes grind to others’ polished highlights.  

**Result?** Exhaustion. Doubt. That voice whispering, *“This isn’t working.”*  

---

## **The Scheduler Fix: Write Less, Grow More**  

I built the **Substack Notes Scheduler** to solve this.  

Here’s how it works:  
- **Batch your writing**: Dedicate *one* weekend day to 2–3 deep-writing sessions.  
- **Plan ahead**: Draft 3–4 notes in a single flow, then schedule them.  
- **Protect your energy**: No more “What should I write *today*?” panic.  

**Why it works**:  
- You trade scattered effort for focused creativity.  
- You build momentum without burning out.  
- You let your best ideas *breathe* before publishing.  

---

## **3 Anti-Burnout Rules to Steal**  

### 1. **Set a “Minimum Viable Note”**  
Don’t aim for perfection. Define “good enough”:  
- 400 words? A strong hook + one actionable tip?  
- A raw, honest story?  
**Done is better than perfect.**  

### 2. **Track Wins, Not Just Words**  
Burnout thrives in the dark. Shine a light:  
- Did someone comment? Save that message.  
- Did a reader DM you? Print it out.  
- Did you *finish* a note? Celebrate.  

### 3. **Rest is Strategic**  
Take a full day off *every week*. No writing. No scrolling.  
Your brain needs space to reset. Your audience needs *time* to engage.  

---

## **Your Next Step**  

Stop grinding. Start *orchestrating*.  

Try the Substack Notes Scheduler this weekend:  
1. Block 3 hours for writing.  
2. Draft 2 notes (yes, *two*).  
3. Schedule them for next week.  

You won’t burn out by writing smarter.  

**What if you tried this tomorrow?**  

---  

*P.S. Growth isn’t about pushing harder. It’s about pulling back, then leaning in.*

Evaluation and critique by an “expert Substack growth coach”

In this section, AI will take the role of an expert Substack growth coach hired to REVIEW a single draft note before publishing. It was requested to evaluate the note strictly from the perspective of the target audience.

AI is scoring the overall publish-readiness of each section of the note, and providing action items will help refine the ideas for the next iteration of the loop.

The Scoring rubric (guideline) I used was:

• 90-100 Publish as-is (minor or no tweaks)

• 75-89 Solid – needs minor edits

• 60-74 Requires improvements before publishing

• <60 Rewrite recommended

The following shows how this supposedly “Viral Note” was scored:

{
  "score": 92,
  "hook": "Strong hook confronts reader's self-blame, immediately relevant to burnout fears",
  "quick_win": "Scheduler solution presented clearly as first actionable answer to burnout",
  "body": "Delivers value through concrete batching strategy and anti-burnout rules within 150 words",
  "cta": "Clear weekend challenge ends with provocative question to spark action",
  "emotional_resonance": "Effectively addresses pain of burnout and desire for sustainable growth",
  "action_items": [
    "Add a specific example of a 'Minimum Viable Note' to make Rule #1 more tangible",
    "Consider adding a brief success testimonial or data point about scheduler effectiveness",
    "Clarify exactly how the scheduler tool works in practice for first-time users"
  ]
}

The real audience test

This very small model (qwen3:32b) has only 32.8B parameters. It runs locally on my MacBook Pro and takes quite a while to process this data.

It doesn’t perform anywhere near the level of the best humans, or at the level of the best commercial AI models that are 1000x larger.

To close the learning loop, I would need to push this out and let the audience determine what this note means to them. I already have the tools to pull the engagement metrics automatically, so adding that to the Python script would be a trivial exercise.

I suppose the real question here is: should I build an automated Note Factory like this?

My MacBook could churn out 110 Notes a day for $0.55 in electricity.

At a 4% conversion rate? That’s $127 in revenue every day.

Welcome to the AI content mines: low cost, high volume, no soul.

Why does this matter?

Learning loops aren’t just a mental model - they’re a roadmap.

• For humans: It’s a reminder that growth comes from iteration, not perfection. Feedback isn’t failure but fuel for improvement.

• For AI: It’s a challenge to build systems that don’t just process data, but learn in context, like us. My tiny experiment above took just half a day to create, using AI, of course.

• For creators and builders: The loop is how we ship, adjust, and ship again.

In a world obsessed with fast results, the real edge is knowing how to loop better than anyone else.

Closing the loop

Here’s the big idea:

An AI without a learning loop is like a stand-up comic who never hears the laughter.

It talks. But it doesn’t grow.

Whether you’re writing, coding, parenting, or just figuring stuff out:

Loop. Iterate. Evolve.

That’s where the good stuff lives.

Want more stories on learning, creativity, and intelligent tools?
👉 Subscribe for weekly loops in your inbox.