I’ve just published a new template! I’ve been doubling down on Instagram since my last channel removal (yeah… it happened again). I learned a few new tricks from that experience. But long story short — if the worst were to happen to my Instagram account, what then?

I realized I needed a backup system, something that would save me a lot of effort. I had to manually reupload everything and then manually search for the data again based on the titles… such a pain. If that ever happens again, I’m automating the whole process.

Right now, I feel like I’m 0% focused on content and 100% focused on survival. Luckily, my workflow is relatively streamlined, so I can get things done without too much effort. Still, I haven’t been able to fully catch up or complete a proper restore (at least not as much as possible).

But enough rambling!

How it works

This workflow provides a complete backup solution for your Instagram video content with intelligent caption parsing:

• Fetches your Instagram account ID and videos (VIDEO and REELS types)
• Parses captions into structured fields:

Title: Everything before the first hashtag
Description: Everything after the first hashtag (includes all tags)
Tag List: All hashtags extracted as an array
Description Full: Complete original caption text

• Downloads videos in maximum available quality from Instagram
• Uploads videos to a designated Google Drive folder
• Creates/updates a JSON metadata file with all video details
• Prevents duplicates using n8n Data Tables with account-level filtering

Closing Thoughts

I’m glad that I made this, I’m still exploring a similar YT solution, I have a draft but I’ll continue working on it.

I made quite a few things here and there, but all the efforts are quite distributed, so I haven’t been able to write as much as I would it wanted it.

Ah before I forget, the missing part for this workflow would be an improved version with the cover image downloaded as well.

And that’s for another time!

I had on my backlog having an alternative to share posts without having to make a Medium account to read them, not to mention other restrictions like the number of stories. So while reading, I saw two approaches to this issue:

• Paid tiers: The writer provides a Friend link that allows you to read the post for free.
• Substack alternative: The Writer provides a link to Substack where it’s the same post with almost the same features.

Another option I was working on is having your posts on a dedicated site by extracting them from Medium through the RapidAPI.

The Repository has everything to run, there are free API calls per month, so it’s quite useful. Here is the post explaining how to run it.

The problem with this approach is that you have to maintain the site, maybe buy a nice domain, so even with AI tools it’s extra work. I’ll keep it since I like having backups more than ever (Thanks YouTube). That also reminds me, I made an n8n workflow to backup Instagram videos.

Going back to Substack…

I made an account at the beginning of the year, but when trying to copy-paste the content, things went south.

Fast forward to today: I saw there was an API for Substack, and since I’m already backing up posts… I could do a cool workflow that uploads them after doing the backup.

And that’s my current adventure. I requested access to the API, so we’ll see how it plays out.

Still, Medium will be my main platform since it’s comfortable and I haven’t found a way to put multiple images in line in, or I haven’t tried hard enough on Substack, but time will tell.

Peace!

Shout out to YouTube, for teaching me that platforms can just… decide things.

TL;DR: code here.

I’ve been writing on Medium for a while now, and while I like the editor and the reach, I kept thinking: what if I want to share something without asking people to make an account? What if Medium changes their paywall rules again? What if I just want my own backup?

So I made a simple script to automatically fetch Medium posts, convert them to Markdown, download the images, and host them on my own static site. For free. With full control.

The Problem

Medium has some friction points:

• Free article limits: Readers hit a wall after 3–5 stories per month
• Account requirements: Not everyone wants to sign up
• Paywall changes: Policies shift, and your content gets caught in it
• No real backup: Export exists, but it’s manual and messy

I wanted an alternative that’s accessible and automatic.

The Solution: Medium API (RapidAPI) + Astro

Here’s the flow:

Medium Platform
    ↓ (RapidAPI)
Python script fetches posts
    ↓
Converts to Markdown + downloads images
    ↓
Syncs to Astro static site
    ↓
Deploys to GitHub Pages (free hosting)

What You Need

• Medium API via RapidAPI: Free tier gives you 150 calls/month (enough for testing)
• Python: For the fetch script

How It Works

1. Fetch posts from Medium using their API
2. Convert to Markdown with full metadata (title, date, tags)
3. Download images locally so you’re not dependent on Medium’s CDN
4. Sync to Astro which generates a fast static site
5. Deploy to GitHub Pages for free hosting

The whole thing runs with make fetch && make build. That's it.

Why This Matters

• Content ownership: My writing stays mine, even if Medium changes policies or pricing.
• No paywalls: Readers can access everything without hitting limits.
• Speed: Static sites load instantly. No Medium bloat.
• Custom design: I can make it look however I want (still working on that part).
• Backups: Git version control means nothing gets lost.

The Setup (Quick Version)

Download the repository and execute:

# 1. Install dependencies
pip install medium-api
cd site && npm install

# 2. Get a RapidAPI key (free tier works)
export RAPIDAPI_KEY="your-key"

# 3. Fetch and build
make fetch
make build

# 4. Preview
make preview

Visit http://localhost:4321/ and your posts are there.

To deploy, just push to GitHub and enable GitHub Pages. Free hosting, done.

Costs

The free tier (150 API calls/month) is enough for monthly backups. Each fetch uses 1–3 calls per post, so if you have 30 posts, that’s ~90 calls. You’re good.

If you want daily syncs or automation, the Pro plan is $4.99/month for 2,500 calls. Still way cheaper than paying for hosting elsewhere.

What’s Next

This is just backup and hosting, but the Medium API can do way more:

• Track engagement over time (claps, responses)
• Aggregate content from specific publications
• Build custom analytics dashboards
• Cross-post to Dev.to, Hashnode, WordPress automatically
• Research trends and topics at scale

I’m exploring some of those, but for now I just wanted my posts in a place I control. Mission accomplished.

Resources

• Medium API Documentation
• RapidAPI Hub
• Astro Docs

And that’s it. Now I have backups, no paywalls, and peace of mind.

Time will tell if I actually maintain this or if it just becomes another abandoned project, but at least the automation is there.

Cursor is great inside the editor. Claude Code is that same vibe, but in your terminal, and it can spin up parallel subagents to knock out tasks for you.

This post is Part 1 (overview). Part 2 is the one with the “stop wasting tokens” habits and workflows, based on how I burned through my weekly limit in 3 days.

What Claude Code Is

Claude Code is an AI coding assistant you run from the command line. You start it, describe a task, and it works inside your project folder.

The big difference vs plain chat is that Claude Code can:

• See your repo (read files you point it to)
• Make changes (edits, new files)
• Run commands (tests, builds, git commands, etc.)
• Pause and ask permission before it does anything tool-y (run commands, edit files, etc.)
• Use subagents in parallel when you ask it to split work into multiple threads

So yeah, it’s interactive. It’s not just “answering” you, it’s collaborating with you.

Should You Subscribe To It?

This is the part nobody talks about until the bill shows up.

Depending on your setup, you might be already paying for:

• ChatGPT (subscription) for general use as chat
• Cursor (subscription) for IDE help

And now:

• Claude Code (subscription or API usage) for terminal work

That’s… a lot. And yeah, sometimes it’s redundant, because the model overlap is real.

What You’re Actually Paying For

Why does the same prompt sometimes work better in one tool than another?

Because you’re not only sending “your prompt.” Each tool silently changes the situation around the model:

• Different system prompts: tools add their own instructions (style, safety, what actions are allowed, how to format edits, etc.).
• Different context injection: the tool might automatically include repo context, open files, diffs, project memory, or nothing.
• Different retrieval: Cursor can pull relevant snippets via indexing/search; Claude chat can reuse project knowledge; Claude Code can read real files on demand (and it often has CLAUDE.md guidance).
• Different tool access: Claude Code can run commands and edit files (with permissions). That changes how the model “thinks” about solving the task, because it can verify by running tests instead of guessing.
• Different guardrails/permissions: the model may be “more cautious” or more incremental depending on what the wrapper allows without asking you 20 times.

So no, you’re usually not paying for a magical new AI. You’re paying for a different environment around the same AI.

Here’s the practical difference in human terms:

• Cursor: best for high-speed IDE work — search + edit + refactor loops — because it keeps context tight and iteration cheap.
• ChatGPT (App/Web): best for general Q&A and writing when you want the “just chat” experience (and often a different feel vs the same model embedded inside an IDE).
• Claude (Desktop/Web): best when the job is thinking, writing, or discussing. It’s a great chat UX, and projects/memory can help for repeated context.
• Claude Code: best when the job is “touch multiple files + run commands + verify.” It can act more like a teammate because it can actually do things (tests, builds, git diff, etc.), not just suggest them.

That’s why the same prompt can feel “smarter” in one place: not because the model changed, but because the tool fed it better context and gave it better ways to verify.

What I Personally Optimize for

• If I need fast coding in one file: Cursor.
• If I need chat + brainstorming + writing: ChatGPT and Cursor.
• If I need repo-level work + running commands + taking initiative and actions + multi-file changes: Claude Code.

So I’m sticking with Cursor for a while, because Claude Code gets me what I want faster, but I burn credits like there’s no tomorrow.

I saw this Claude vs Cursor question a while back, and the best answer (in my opinion) was: “use whatever gets you the work done” 😂

Should You Cut Down On AI?

Claude also has a desktop app (a normal chat app) that, for some people, can replace part of their ChatGPT subscription (depending on what models/features they actually use).

It also has IDE integrations (like Visual Studio Code), so in theory that covers the “Cursor approach” too.

I’m not saying “cancel everything right now”. Just that if you’re trying to simplify your stack, it’s an option.

Though… in my experience, free ChatGPT 5.2 in the ChatGPT browser gives me better answers than the same model through Cursor. So I’ll leave that up to interpretation. Maybe it’s just me.

Cost Basics

Claude Code can be used in two different “billing styles,” and it changes how you should think about cost:

• Subscription (Pro/Max): you’re not paying per message. You get a usage allowance, and the important part is the limits (you can run out).
• API-style (pay per token): you’re literally billed by tokens, like the Claude API. Bigger context = more tokens = more cost.

Practical notes:

• Every file it reads costs tokens/usage
• Long conversations cost more over time
• Big “scan my whole repo” requests are expensive
• On subscription, the commands that matter are usually /usage (limits) and /stats (patterns)

The Real Limits

Claude has two big buckets of limits:

• 5‑hour session limit: you get an allowance, you burn through it, then it resets every ~5 hours.
• Weekly limits: a cap across the week that can hard-stop you until the reset.

Also important: these limits apply across Claude AND Claude Code. So if you spend the day chatting on the desktop/web app, your terminal sessions feel it (and vice versa).

And yes, the “soft limit” feeling is real: you can hit it fast if you go into backlog mode (I hit it in ~20 minutes once). The weekly limit is worse. I hit that too and got four days without Claude. Pain.

One more detail that matters: the Usage page shows weekly limits separately for Opus vs all other models, so you can accidentally burn the “expensive bucket” first.

How to See Your Limits (So You Don’t Guess)

Claude will warn you when you’re getting close, but those warnings can be *fast* (and easy to miss if you’re in the zone).

So I treat limit-checking like checking the gas tank. You have two good options:

In the Claude app (best UI): Settings → Usage https://claude.ai/settings/usage

This shows progress bars for:

• Current session (how much you used + time remaining until the 5‑hour reset)
• Weekly limits (with reset timing; usually split for Opus vs other models)

In Claude Code (terminal): run /usage

If you’re doing a long session, just make it a habit to run /usage every once in a while (or right before you start a big “ok now refactor the whole thing” request).

(Part 2 is where I show the habits that keep this under control.)

Setup: the two commands you need

Start Claude Code

From your project folder: /claude

Generate a starter CLAUDE.md (do this once)

Inside Claude Code, run: /init

That will generate a starter CLAUDE.md so Claude doesn’t have to re-learn “how to run tests” and “what not to touch” every session.

What’s in Part 2

• The right way to use sessions (claude -c vs claude -r)
• How permissions work (and how not to click “Allow” 200 times)
• Parallel subagents (“do these 3 things at once”)
• The token-saving workflow I actually use

If you want the practical part, go to Part 2 (Eventually).

Peace!

Me letting the AI take the control:

I had a recent realization: making people “shine,” as one of my clients put it, isn’t actually that common.

And I think I learned how to do it playing Yu-Gi-Oh in high school.

So no, I wasn’t wasting my life back then.

…Probably.

The Ender’s Game Baseline

It’s not related to Yu-Gi-Oh! but I’ve internalized a quote from the Ender’s Game books. It goes something like: “You are not better than your general.”

(I haven’t been able to pin down the quote, but probably it should be from Shadow saga, Bean’s perspective. But I digress.)

The point is: you can be a great employee, but if your boss is driving toward a cliff, there’s only so much you can do.

I value having the agency to decide how I handle tasks and projects. My aim is to simplify my work and the work of whoever I report to. And if I know the way forward, I’ll make sure we take the optimal path.

Development Arc

I spent a lot of time on Yu-Gi-Oh. Looking back, like most things, it was a side quest to acquire something else. In this case, it was the mentality.

At that age, most of my skills weren’t developed at all. I’d say I only started mentally coming online around 22, and by 25 I had real potency and leverage.

That’s why I don’t blame Fusion users. If you start with raw power, the optimal move is often to enter the field and get things done. Back then, for me, it was about working with scraps, assembling whatever I could, and finding a way to stay in the fight.

Deck Building as Strategy

Coming back to leadership, coordination, and strategy… from my standpoint, it’s not that hard. There were personal challenges, sure. But when it comes to using resources, it’s usually obvious what needs to be played.

What I noticed a few years ago, and again recently when I couldn’t find the words to explain something to someone, is that the movements are so obvious it’s painful to see. I ended up saying:

It’s like, you just need to remove a few copies of that card and the deck will run faster.

For those not fluent in Trading Card Games (TCGs), that’s deck-building advice. But I realized I was unconsciously using that same mental framework.

The idea is simple: build a deck with a good variety of multi-purpose cards, and remove the ones that aren’t worth drawing or are too specific. They might be useful. They might be cool. But most of the time, they’re a poor draw, and a missed chance to play something better.

In some cases, that card even damages your own cards. And then what? Why would you do that to your deck, and to yourself?

Yes. It happens.

People as Cards

When you spend enough time around people, you start to notice patterns. Most have a default way of operating, a predictable set of strengths, and habits that repeat a little too often.
As you interact more, you discover additional skills, preferences, and above all, conditions under which they work best.

Assigning someone a task that’s far beyond their current scope and requires excessive setup is rarely a smart move. But when the conditions are right and the situation fits their strengths, things click. In those moments, people can contribute far more than they initially seem capable of.

Even when you can’t unlock new skills right away, positioning still matters. The way people complement each other, the timing of their involvement, and how responsibilities are combined can make a significant difference.

Sometimes you don’t get to choose the team or the timing. But if you can plan a few steps ahead, you can turn a difficult situation around by preparing the ground carefully.

That’s the mentality behind Synchro Summoning.

Summoning Philosophies

I’ve never liked the mentality behind Fusion Summoning. To me, it feels like handholding. Impactful, sure, but most of the time it’s a waste of resources and a loss of versatility.
In practical terms, it’s like a client expecting three Blue-Eyes White Dragons by fusing a high-level senior with a semi-senior. It might work, but you end up with a single outcome instead of two or three, after burning three resources to get there.

Now that I think about it, the real corporate style would be Ritual Summoning. All that structure, process, and ceremony just to get something on the field.

Synchro Summoning works differently. Smaller pieces combine into something bigger, with one element providing alignment and timing. Delta Accel Synchro follows the same logic, just scaled up. Bigger results, same principles.

That’s the style I like. It feels more realistic. Small, well-timed efforts, aligned correctly, can produce outsized results. Not a single flashy move, but multiple configurations that let you respond, adapt, and stay versatile.

Being a Tuner

As a bit of an overshare, I’ve always identified strongly with Quickdraw Synchron. Partly because my deck revolved around it, but also because it became my mental model for speed, level awareness, and timing.

In that sense, I think of myself as a Tuner. Not as the main force, but as the element that helps things align. The role that enables larger outcomes by syncing well with others.

That perspective probably explains how I look at collaboration and teamwork. I focus on how different strengths can line up, how timing matters, and how coordination can unlock results that wouldn’t exist otherwise.

At my current level, that tuning ability can feel overpowered. But it only exists because I didn’t start with most of the foundational skills. I had to build them over time, leaning heavily on my ability to synchronize, adapt, and make things work together while I learned the rest.

The Caveats

Because I tend to think clearly and quickly, there’s an obvious trap for someone like me: over-relying on myself to start every combo. You either burn yourself out, or if you’re managing people, you burn your Tuners.

There are plenty of ways to address that. My current definition is this: it’s like playing with a handicap for fun. You deliberately avoid One Turn Killing everything, let the game breathe, and allow situations to play out. You can still stop the game when you need to, and that’s fine.

In tech, this looks less like going for an OTK and more like hitting damage quotas. Yes, with a good hand you could end the game immediately. But you usually shouldn’t. It’s basically the meme where a junior says a project will take a week, and the senior says three months.

Another adjustment I’ve made is not needing to be the main Tuner anymore. I focus on arranging the table so the game runs smoothly. Worst case, I can always jump in, accelerate things, and tune directly. But that’s the easy way out. There’s a different game I’m enjoying right now, and I’ll leave that story for another time.

To live a long life, some inner peace is required. Battles are fun, but some games are inevitably slow and predictable. That’s often the best moment to let things level up naturally. After a while, if the rewards stop scaling, then you do what’s necessary and nothing more.

In Summary

Yusei is the best Yu-Gi-Oh protagonist, and his ending is a solid illustration of everything I’m talking about here. Yes, getting a real job.

I wouldn’t claim this is the perfect way to think about people or collaboration. But it is an efficient one. It reduces mental overhead, skips unnecessary narrative, and focuses on how different strengths can combine in ways that benefit everyone involved.

At a leadership level, this translates into something simple: don’t force outcomes, don’t burn resources for the sake of speed, and don’t confuse “can” with “should.”

If you’re hitting the quotas, keeping options in hand is often better than ending the game early or overworking yourself. Most of the time, there isn’t much urgency. Some priorities exist just to get things rolling, and once you talk it through, that becomes clear.

And yes,

CARD GAMES ON MOTORCYCLES!

Peace!

Most existing workflows use spreadsheets to handle scheduling. That’s too cumbersome for me. I opted for a simpler approach: run at a certain time, pull a certain amount of videos, then add a delay between uploads so they’re spaced out with a reasonable gap.

If you’re new to the Instagram API setup, check Part 1 first.

What This Workflow Does

1. Runs daily (or whenever you set it)
2. Fetches your latest Instagram posts
3. Filters to only videos (Reels included)
4. Checks a Data Table so it doesn’t re-upload
5. Waits between uploads (configurable gap)
6. Extracts hashtags as YouTube tags
7. Uploads to YouTube with smart title truncation
8. Saves the record

Simple. Boring. Reliable.

The Configuration Node

Everything lives in one JSON blob:

{
  "includeSourceLink": false,
  "waitTimeoutSeconds": 900,
  "maxTitleLength": 100,
  "categoryId": "22",
  "privacyStatus": "public",
  "notifySubscribers": false,
  "defaultLanguage": "en",
  "ageRestricted": true
}

Key settings:

• includeSourceLink: Adds the Instagram permalink to the YouTube description. Set false if your channel isn't verified (external links get flagged).
• waitTimeoutSeconds: Gap between uploads. 900 = 15 minutes. YouTube doesn’t love rapid-fire uploads.
• ageRestricted: Set true for 18+ content. This is the "made for kids" COPPA setting.
• categoryId: YouTube category. 22 = People & Blogs. See the table below.

Category IDs

• 22: People & Blogs
• 24: Entertainment
• 20: Gaming
• 10: Music

Setup Steps

1. Create the Data Table

In n8n: Data Tables → create one with two columns:

• postId (string) ← Instagram media ID
• youtubeId (string) ← YouTube video ID after upload

This is your “already uploaded” memory.

2. Set Your Instagram Credentials

In the Fetch Instagram Posts node, set your Bearer Auth credential (same token from Part 1).

3. Set Your YouTube Credentials

In the Upload to YouTube node, connect your YouTube OAuth2 credential.

If you haven’t set up YouTube OAuth in n8n before: it’s the standard Google OAuth dance. Create credentials in Google Cloud Console, enable the YouTube Data API v3, and add the OAuth consent screen.

4. Import and Activate

Paste the JSON into a new workflow, update the credentials, save, and activate.

How the Title Works

The Code node does the heavy lifting:

1. Grabs the caption
2. Extracts text before the first hashtag → that’s your title
3. If no text before hashtags, uses the first hashtag as the title
4. Truncates smartly at word boundaries (no cutting mid-word)
5. Pulls all hashtags as YouTube tags (max 10, cleaned up)

So if your Instagram caption is:

Check out this figure unboxing! #figma #anime #unboxing

The YouTube title becomes: Check out this figure unboxing!

And the tags become: #figma #anime #unboxing

Why the Wait Node?

To space out the publishing of videos. The workflow processes one video, waits (default 15 min), then moves to the next.

For testing, just use 1 second.

If you have 10 videos to upload, that’s ~2.5 hours to upload them all. Slow but natural.

Common Gotchas

• “External links not allowed”: Unverified channels can’t add external links in descriptions. Set includeSourceLink: false.
• OAuth token expired: YouTube OAuth tokens expire. Re-authenticate in n8n if uploads fail.
• Video too long: Instagram Reels are usually under 90 seconds, so this rarely matters. But YouTube has its own limits for unverified channels.
• Wrong category: If your content keeps getting flagged, try a different categoryId.

The Workflow JSON

I submitted it to n8n as a template but until I get some feedback on it. Here’s the JSON in my n8n-templates repo:

• Instagram to YouTube Workflow

Edit: the template was approved, available on the n8n page.

What’s Next

Trends… the last boss, at least of this season.

Resources

• Part 1: Getting Instagram Posts with n8n
• Part 2: Posting to Twitter/X
• Adding Multi-Accounts
• YouTube Data API v3
• n8n YouTube Node Docs

So I’ve been playing with AI image generation for a while now. And I’ve been facing something really annoying more and more: the more I iterated over an image with prompts, the worse it looked.

It was obvious that it was the AI and the iterations causing it, but I didn’t know the name. And I didn’t put much effort into googling beyond the first few results. Most of my searches just gave me stuff about hallucinations, which wasn’t quite it.

So I did something different today. I asked Google AI for an answer rather than for an image (for a change of pace). And I finally got the name for this issue.

The Problem: Generational Loss

Turns out, this has a name borrowed from analog media: Generational Loss (or Iterative Degradation).

Think of it like making a photocopy of a photocopy of a photocopy. Each copy loses a bit of quality. Eventually, you end up with a blurry mess.

In AI terms, here’s what’s happening:

1. The AI analyzes your image pixels
2. It compresses them into mathematical noise (latent space)
3. It re-imagines what that noise represents based on its training data
4. It outputs a “new” version

Every time it does this, it makes small guesses. Small errors. And when you feed that result back in, the AI mistakes its previous errors for intentional details and amplifies them. That’s why your colors get saturated, your lines get thick, and everything starts looking “overdrawn.”

Even though I’m using the Nanobana Pro model (which I really like), it still suffers from the same problem. And since I really like to refine and try things out, I run into this a lot.

Why “Clean This Up” Makes It Worse

When you prompt the AI to “clear colors” or “make it look better,” it often interprets this by:

• Cranking up contrast and saturation to make it “pop”
• Thickening lines to make them appear “sharper”

Which is… the opposite of what you wanted. The AI has a different definition of “better.”

And there you have it: a correct answer but poor results.

Better Prompts for Fixing Overdrawn and Oversaturated Images

To actually fix this, you can include specific stylistic instructions that steer the AI toward softness, precision, and balance.

Based on my current case, it would be something like this:

Approach 1: The “Soft and Clean” Approach

Best for: Reducing saturation

Prompt: A high-quality chibi illustration of this character. Soft pastel colors,
balanced saturation, clean and thin line art, smooth gradients, gentle
shading. No harsh contrast.

Why it works: Words like “soft,” “pastel,” and “gentle” directly counter high saturation. “Thin line art” fixes the overdrawn look.

Approach 2: The “Professional Vector” Approach

Best for: Crispness without mess

Prompt: Clean vector art style illustration of the character. Crisp and precise
linework, flat colors with minimal cel-shading, bright but balanced color
palette. Professional digital art, sharp focus.

Why it works: “Vector art” implies clean, mathematical precision rather than sketchy lines. “Flat colors” stops the AI from adding confusing shadow layers.

Approach 3: The Negative Prompt Approach

Best for: When your AI tool supports it

Positive Prompt: Cleaned up high-resolution anime illustration of the character.

Negative Prompt: Oversaturated, high contrast, thick black lines, muddy colors, blurry,
sketchy, harsh shadows.

Why it works: You’re specifically forbidding the elements that caused the problem in the first place.

Recovering Already-Generated Images

If you’ve already gone through a few iterations and your image looks rough, here are some prompts to try to recover it:

For oversaturated colors:

Reduce saturation, soft and muted color palette, natural lighting, 
no color boosting.

For thick/messy lines:

Refine linework, thin clean lines, precise outlines, reduce line weight, 
smooth edges.

For general “overcooked” look:

Subtle refinements only, preserve original style, minimal changes, 
gentle cleanup, maintain original color balance.

The key is being explicit about what you DON’T want to change, not just what you want to fix.

TL;DR

• The problem: Feeding AI images back into AI causes quality loss (Generational Loss)
• Why: The AI mistakes its own errors for intentional details and amplifies them
• Fix: Use specific prompts like “soft pastel colors” and “thin line art”
• Avoid: Vague prompts like “make it better” or “clean this up”
• Negative prompts: Explicitly ban “oversaturated, thick lines, harsh contrast”

And that’s it! Hopefully this saves you some frustration. Also, this relates to the concept of Model Collapse, which is basically inbred models through models trained primarily on AI images. Something to think about.

Peace! ✌️

In this post I go through the “creative process” which is prompting until I realize what I want and then try to do it, while keeping my sanity points at a decent threshold (as always).

As with previous AI posts, I’ll include some of the prompts used in quotes.

TL;DR: I realized that I just wanted the balloon while keeping the suit. And that I had to question what I really wanted instead of prompting until I hit gold. Feasible if you don’t pay for credits and use the free tier for Nanobana Pro. Otherwise, think a little, make some sketches or something.

The result was a logo recognizable to old subs but simplified and more catchy.

Context

When I finally came to terms with the YouTube account deletion, after reuploading and holding off on things for a while, I needed to do something for my “main” logo. It was a little awkward to have it there, looking at me in disappointment. “How did they get you?” You know, the equivalent of a woman cutting her hair to represent transformation.

I had a random logo that I generated for Medium amidst all of this: a smiley balloon with a red cross on the face. It was for the “cancellable” aspect of it rather than “no happiness,” though at this point I don’t think it’s that different.

Since that version of me is dead, it was odd doing more things with the old logo. I wanted something different.

I wanted to keep the new logo, so I tried a few things.

About the New YouTube Channel Logos

This will be for another post, but I divided the content from the original channel into five channels at the moment of writing this.

That will be its own post. The takeaway from this is that I finally used the mascot character (Chibi Ver.) for the YouTube logos, multiple colors and ordered as Power Rangers.

First Interation: Shooting in the Dark

Since my YouTube accounts had different colors, I wanted black for the main one even if it’s just an “index” for the rest of the channels. But I wanted to keep the current Medium logo.

Prompt: I want a perfect square but circle inside with abstract art representation of the character focused on colors.

Even tried the male version for the funzies. Not sharing because it’s mid.

In my mind I wanted something akin to shapes, simplified, that also had the X. So I wasn’t sure how or where to go.

Don’t know how it went to Gryffindor but well. Back to the balloon I guess.

Second Iteration: Trying to Enhance the Current Logo

Prompt: Make it better to be in a square image and being mostly used in circles for profile picture.

That didn’t work as expected, so I passed the base logo and requested to have it more like that, bright colors, and an X for the balloon.

Prompt: maybe try a more catchy good combination of colors but I like where it’s going, also, maybe make the X follow the contour of the balloon.

Yes, the prompt “make the cross part of the balloon” backfired horribly.

Fourth Iteration: Balloon

Prompt: Replace the head with a balloon with an X on the head and a happy expression.

So an X looks fine as eyes, so two may work, then let’s try it.

Prompt: make two X to be the eyes.

Prompt: don’t forget the smile.

Prompt: make the X the same color of the tie and add the rest of the suit on the circle (the black part).

Prompt: keep the suit red

Prompt: improve the quality, remove noise and move the mouth to the correct position

I tried other iterations since I was happy but not conformed with the results…

Zoomed out and realized that it’s a great logo if you aim for serial killer. After a while it reminded me of the Red John smiley face.

Fifth Iteration: Less Creepy

Prompt: can you make the X white and change the balloon color to an orange like the tie, and change the tie color to one that makes sense.

The blue makes sense, represents Chilean flag colors, so great job AI, and bad me because I didn’t notice nor care until I saw it there.

Prompt: Please increase the size of the balloon person within the circle so it appears on top of the border.

Part 6: Vento Aureo

This had more to do with the borders. Since my last prompt asked to move the head more to the top, the upper border will be small, which I’m fine with.

The issue with the border colors was noticeable if you added it to pages that didn’t allow you to adjust the position of the image. For those cases, you would see white/empty space.

So to avoid that stupid problem, I:

1. Locally removed the background with the Image Preview from macOS.
2. Upscaled it with Upscale Media.
3. Shared it to my phone.
4. Opened Lightroom, adjusted colors, removed the thingy from the circle.
5. Shared it back to my computer.
6. Edited it on Canva to add the background.

And finally it was ready to share on social media. And since it was Friday, with this, it’s more than enough.

Results

Quite happy with the result. I’m really not there, since my sub accounts have the content, but the old me is there. It kinda works both in a spiritual and organizational way.

I believe that, and that makes me happy.

What’s Next

If there’s something to take from this, it’s that prompting until you hit gold is inefficient. Stopping to ask yourself “what do I actually want?” saves time. And if you’re not on a free tier burning through credits, maybe sketch something first.

That said, editing skills and knowing the right tools go a long way. The AI got me 80% there, but Canva, Lightroom, and macOS Image Preview handled the rest.

You don’t need to be a designer, but knowing how to remove a background or adjust colors makes a difference.

Next post (maybe not from this series) will be about organization. Keeping your online presence consistent requires planning, and things scale fast. Especially when you’re juggling multiple accounts and activities. But that will be for another time.

Bonus

It’s been a while since I tried to generate something from a quick draft. Let’s see how it goes.

Still too troublesome to make something consistent or exactly as I want without more than 20 minutes of effort.

I’ll try to figure out how to course correct the problem where, after many prompts, it starts saturating the images with lines and weird colors. Like, her face looks more yellow.

Well, that’s for another time.

The Hell Known As YouTube

A series where I share what I’ve learned about growing a YouTube channel. Practical advice, personal takes, and lessons from experience. Some things you’ve probably heard before, others you haven’t.

I hit 1,000 subs about a year after starting my channel, peaked around 1,780, then it died out… I also got a meme channel to the point of being monetizable, but that’s another story.

I’ll keep adding to this post as I remember more. This first iteration is basically what I can recall off the top of my head.

This is a lot of fun, after you go through all the cringe and pain and get used to it. It gets better.

Best Advice

Things that I’ve heard and make sense:

• Do this as a hobby. Making it for a living or full-time will drive you insane.
• The algorithm eventually will hit you, sooner or later. Some people take a few months, others a few years. I believe the average is like 5 years.
• When the algorithm hits, you better have content. Otherwise you won’t retain audience, and also, more chances that they sub. Especially with Shorts.
• Shorts are OP. You can get quick views and more subs with a constant flow.
• Find your niche. It’s easier to connect that way. There are communities, and if your content is something that’s not that common, you have better chances to grow.
• Try to limit your content to one thing and iterate over it until you find your “formula.”
• The “formula” depends on your content. The metric insights can be useful to see what works and what doesn’t. It usually boils down to cutting the fat and keeping a good pace.
• Keep in mind that metrics from video performance will sometimes make you feel miserable.
• It’s sad until you start getting traction and interact with more people. Like streaming to 0 viewers. Having a friend or someone to watch your content will make you cringe a little less.
• The cringe is part of the process.
• The cringe won’t go away, especially when you look at your past videos.
• Perfect is the enemy of good.
• You don’t need to go that hard. The more effort you put in, the more it will hurt when you don’t get the results.
• A video that you pour all your love and skills into may fail, and one with barely any thought put into it may succeed. That’s YouTube, that’s life.
• Time consistency is important, but finding the best time to publish comes first. That’s odd and requires a lot of experimentation. In my case, 12 PM and 3 AM worked great, and certain days were better than others, but after that it was hit or miss.
• Promoting your channel through ads is not worth it unless you are a business.

Impressions

Some things that I noticed with my deceased channel:

• It didn’t have 4K for Shorts until basically half its life. Also, the max duration was 1 minute, when the publicly available info said 3 minutes. My new channels didn’t have that problem.
• Somewhere along the way, timestamps started working. Before that, they didn’t work even with the right pattern.
• I never saw the option to edit Shorts thumbnails. Maybe skill issue.
• People that found my channel said it was hard to find (partially blame the shadow banning), but also that even if the title and description can be translated when viewing, you need to add the additional languages so it’s also searchable that way. And that’s how I made my first n8n template.

Channel

• Keep in mind the settings for your channel. They do impact the visibility and the reach of your videos.
• Make it unique, personalized in your style.

I’ll cover this in another post.

Content

• If you want to have NSFW stuff, don’t push it too hard. Make a secondary account even if it pains you. You will receive fewer strikes and avoid unnecessary problems in the long run.
• Try to go for royalty-free music. If you have any copyrighted content, your video won’t be able to be monetized. I know that because of the meme channel that I mentioned.
• Make good quality content, lol.

Edition

• Be organized.
• Free tier goes a long way. I use CapCut since I didn’t want to try anything else until I felt fully well-versed in it. I’ve learned quite a few tricks, and that’s for another time.
• Have predefined assets and templates that you can reuse. Most editors let you create custom presets that you can copy between items, like sound levels, image adjustments, and so on.
• Thumbnails are important. Canva is a powerful service that goes a long way with the free tier, but you’ll save a lot of time by paying for features you’d otherwise need third-party tools for.
• Storage, especially if you record in 4K, is going to be a problem. Be prepared for that.
• Make backups, even if it’s on another YouTube channel.
• Color grading is horrible if you want the highest resolution. I cover that briefly in this series.
• Again, you don’t need perfection, just flow.
• AI slop will be AI slop.

Quality of Life

• Try to save time in any way possible. Be organized, have shortcuts, and so on, because it gets time consuming. I recall my first videos took me like 8 hours. Then it went down to an hour and a half, by trimming the fat and having default settings ready.
• I’m not fond of intros, which are suggested at certain points to make a channel more “familiar” to recurrent subscribers. But try to make the content standard. Like in my case: spin, overview, quick review, spins to end, outro screen.
• The outro screen that I use is just a dimmed transparent screen to remind viewers to subscribe and like (primo advice), but it reduced the visual clutter.
• Music helps a lot. Just be sure to balance the audio if you’re talking.

What’s Next

Next up, I want to cover customization and the common settings. There’s a lot of info out there about this, but I prefer having it my way and without having to sit through a video for it. The irony I know.

Moving On and Starting Over

My channel got removed, and I don’t want this to be the “I peaked in high school” kind of post. I want to learn and reflect on everything that has happened. Yes, the views, the subscribers, and some videos are gone, but not everything is lost.

It’s horrible. Not the channel removal itself, or that it’s almost starting all over again. It’s going through the first months of a new channel again. Slowwwww growth, but this time 5x worse.

I remember the first month, like one constant viewer 😂

Moe: Ahora toquen, el público se impacienta.
Barney: ¡Queremos a los Chili Peppers! ¡Queremos a los Chili Peppers!

Being active in communities can help, but otherwise you depend a lot on the algorithm. That’s something I’ll cover later.

What Hurts the Most

What hurts is that I have to go through my old videos, try not to cringe, and upload them again. Mainly 1/6 dolls and anime figure unboxings, which by my current standards (or depending on my mood), are not good enough. But they may be useful for others.

I aimed for stuff that was less likely to be covered and did it in my own style. That was the original reason for the channel: to see what I couldn’t find elsewhere, without having to buy things just to find out if they were compatible between different 1/6 scale brands.

What Helped Me Move On

Some people cared about it. They reached out and asked if I was going to come back. They said my content was useful for them. And that means the world, even if I had to cringe my way through re-uploading.

I started because I wanted to make the videos that I wanted to see, even if that risked the account. All while improving my photography, videography, and editing skills.

And it’s fun!

Why This Series Exists

Since I’m starting again on YouTube, I have some things to share.

I hit 1,000 subs about a year after creating the channel for dolls and anime figures, and it died out around 1,780 a few months later. So I have a few things that may be useful. Some are the same advice you find everywhere, plus my personal takes. Not the best performance overall, but decent experience compared to others.

Funny enough, I wasn’t able to monetize it. I couldn’t hit the quota of watch hours. I thought I reached 4,000+ and even received the notification, but YouTube delays and reports only the hours they consider valid. For reference, mine was actually around 2,900.

Doing what’s considered “not technically 18+ but 18+” content is a little complicated. In most cases you won’t be able to monetize regardless, since that type of content limits the advertisers, results in less revenue, and reduces views if you get shadow banned or strikes. That’s why some channels that could be more explicit choose not to be. Sounds like cope, but that’s why I stopped caring about monetization and focused on improving my skills. I take pride in that, though looking back, I cringe a lot.

But that’s not all. I also have another channel where I occasionally upload references and memes to use at work. That one actually hit the YouTube milestones to be monetized, since it had better performance and some videos hit the right trends to get pushed by the algorithm.

And guess what? It got rejected since it’s considered “low effort content.” Fair enough. But I was able to glimpse the pains and joys of what it looks like to be monetizable. You can try again after a certain period, but I ain’t going to give more love than necessary to that channel.

TL;DR: That’s why meme/anime channels start doing weird or longer content after a while and replace music. You can’t live off the small stuff and copyright will negate all earnings if applies.

What’s Next

Next up, I want to cover the best advice I’ve picked up, impressions from my main channel, content tips, and quality of life stuff. I’ll keep adding to this series as I remember more.