The Human Proof of Understanding Settings: How a Non-Engineer Self-Destructed by Thinking 'It Means the Same Thing'

It means the same thing, so it should be fine, I'm going to tell you the story of how everything ended the moment I hit Enter.

This isn't a case of 'the tool is bad' or 'the AI is stupid'.

It's a story of my own making, where I built something by dumping everything on AI without understanding a single line of the settings.

In my February 21st article, I wrote about 'not diving into customizations immediately when using a new tool'.

That article was about 'tool issues'.

This time is different.

This time, it's a story about a much worse, self-inflicted failure: 'not understanding the meaning of the settings I dumped onto the AI'.

This is a record of my human proof.

Chapter 1: The Late-Night 'Usage Limit Reached'

2:00 AM.

Only the pale blue light of the PC illuminates the room.

The energy drink can is empty.

And the merciless text floating on the screen—

"Usage limit reached"

"I'm so naive... I'm so soft!"

I tap the calculator.

In my workflow, where I set up an agent team and pass them my past articles, thought texts, and research data in their entirety every time, I consume over 5% of my usage limit for a single article generation.

By simple calculation, I hit the limit after writing 20 articles.

While keeping up with my daily posting routine, this usage issue was a ticking time bomb.

The limit for Claude Pro is 'a maximum of 45 messages per 5-hour rolling window (effectively 10-40)'.

Internally, a single 'simple file edit' triggers 8-12 API calls, and as the context builds up, there are cases where a single final command exceeds 200,000 tokens.

'Leave everything to the AI agent for full automation'

I, too, had believed the sleep-talking of such pseudo-entrepreneurs somewhere along the line.

Like the concrete at Kamurocho Hills, it looked smooth and clean on the surface but was hollow inside—that was the way I had been using it.

The hustle of usage limits is not sweet.

From here, my counterattack began.

Chapter 2: The Counterattack of Obsidian x Claude Code x Gemini Embedding 2

As a solution to the cost problem, I chose to introduce Obsidian as an 'external memory device (exoskeleton)'.

If I don't have to pass the entire massive context every time, usage will drop drastically.

Store notes in an Obsidian Vault, vectorize them with Gemini Embedding 2, and have Claude Code retrieve only the necessary information via RAG search.

If I can do this, my thoughts from 80 past articles can be passed to Claude 'only as much as needed, when needed'.

I used Codex (ChatGPT 5.4) to create the design documents.

It's a division of labor where I solidify the logical design with Codex and have Claude Code implement it.

[Supplement] Addressing the contradiction with the 3/20 article head-on

In the 3/20 article, I wrote, 'Use Antigravity as a designer'.

This time, I'm using Codex. It's a contradiction, right?

Yes, it is. But I have an answer ready.

Antigravity excels at implementation-oriented design including parallel processing and browser testing, while Codex is suited for pure logical design and theoretical construction.

For the task of building design documents from scratch, ChatGPT 5.4 is the strongest, and it has a free tier.

Tools should be used for their intended purpose. If there's a top-tier free tier available, it's a site manager's call to use that first.

The implementation proceeded in stages.

Writing Python scripts, resolving API errors with official code, and fixing the note-loading part of Obsidian.

Since one-shot success is basically non-existent, this level of fixing is a breeze.

——That said, there was also the time I spent two hours copy-pasting in Antigravity before realizing it could be solved in seconds.

I wasted two hours trying to migrate my note articles manually.

While copy-pasting, thinking, 'I'll fix this part of this article... then this one...', I pushed forward with the naive estimate that 'Well, it should take 2-3 hours'.

Whoa!!!
This is way too tedious!!!

When I typed a single line into Antigravity saying, 'Copy all my articles from this site, save them as .md files, and put them in this folder,' it was finished in 5 to 10 minutes.

What on earth was I doing for two hours with bloodshot eyes!

And then, the relief that 'it's finished' arrived.

It's done. Now all that's left is to hit the launch trigger.

That relief was the beginning of everything.

Chapter 3: The Day I Lost Everything with the Final Enter (The Core of the Double Failure)

Let me insert a story from when I was working as a cook.

Learning how to hold a kitchen knife and understanding what the knife is a tool for are two completely different things.

'My hands move. But I don't understand the meaning of the cooking.'—That state will inevitably reach a limit at some point.

The AI configuration file was the same.

'I can write it. But I don't understand the meaning.'—That state certainly existed within me.

The true nature of that is a story for a little later.

——It was time to input the activation trigger for the agent team.

“Turn this text into a Note article.”

This was the activation trigger specified by the AI.

It was a phrase that had to be followed word for word.

This is what I typed in.

“Turn the following manuscript into a note article in I-Kit (Majima dialect).”

The meaning is the same, right?

——Thinking that, I hit Enter.

The agent team did not activate.

Claude returned an article that felt like 'cold takoyaki' all on its own.

It had no body heat. It had no sense of presence. No demon editor-in-chief, no investigator, no writer.

The moment I thought, 'The meaning is the same,' the AI moved along a completely different path.

Rule-based systems operate on 'symbols and conditions,' not 'context.'

'The same in spirit' doesn't work.

Hey, you.

Haven't you ever typed in a trigger using a loose translation, thinking, 'It's fine if the meaning is the same'?

——I'm sure you have. I was the one who did it. And it didn't end there.

I reached my usage limit.

I couldn't write the article for that day.

Thinking, 'It will work if I just fix the activation trigger issue,' I went to sleep.

At this point, I still hadn't realized the real problem.

Chapter 4: The True Sin of Using Settings Without Knowing What They Mean

The true cause became clear in the chat the next day.

I should have just typed the trigger phrase for the launch accurately

No. That's not it.

There was one line in SKILL.md.

disable-model-invocation: true

What this meant—I learned for the first time in today's chat.

The moment I saw the screen, the strength drained from my back.

My right hand, which was holding a coffee, stopped, and I forgot to even put the cup on the table, remaining like that for a while.

Quietly, slowly, the feeling that 'Ah. I've been wrong the whole time' flowed through my entire body.

.......How is a gritty, middle-aged non-engineer supposed to know that, you idiot!!

This is like me thinking, 'Kiryu-chan will definitely pass through here!'
and, with flared nostrils, hiding inside a giant traffic cone to set up the perfect trap (system), only to end up exhausted from waiting for hours in a sauna-like heat, and when Kiryu-chan finally arrived, I couldn't even put up a decent fight.
It's a ridiculous, stupid end where I confused the means with the ends, just like that afternoon!
I left it all to the AI and set the ultimate trap, but because of my ignorance, I stalled (went bankrupt) before I could even fight!
......Wait, this is just a sad, one-man sumo match caused entirely by my own stupidity!

.......This isn't the time to laugh! Laugh, you idiot!!

disable-model-invocation: true

......So 'disable' meant 'to prohibit'? And 'model-invocation' meant 'to call the model'.
In other words, it meant 'Prohibit model invocation: True'!!

What was going through my head when the AI created the configuration file—

A professional understands the meaning of every line in a configuration file before writing it. That's common sense.

I hadn't done that even one millimeter.

'It's fine because the AI said so'—that's the answer sheet of a fake site supervisor.

You've probably had at least one line in a config file you let an AI write that you approved while thinking, 'I don't really get it, but it's probably fine'—that's it. That is the true nature of the problem.

The true nature of this flag—it's a setting to 'prevent automatic startup by Claude'.

Its intended use is for skills that the user is expected to call directly via `/skill-name`.

It is a flag to 'keep Claude from starting up on its own'.

In other words, this is what happened.

As long as this flag is present, no matter how accurate the startup trigger you provide, Claude will be in a state where it cannot launch the agent team.

That is what I did.

While inputting the perfect startup trigger, I was sealing off the startup with my own hands in the config file.

There isn't even an error message. It fails silently.

It's a living hell where you keep fixing the superficial problem of 'the trigger is bad' without even knowing what the cause is.

There's more.

I thought Plan Mode and Bypass Permission Mode were 'pretty much the same thing'.

They are completely different.

Bypass Permission Mode: 'Expands' the scope of what Claude can execute without permission confirmation (loosening constraints).

Plan Mode: 'Restricts' Claude to read-only and plan creation only (tightening constraints).

The two areconstraints in opposite directions.

Thinking they were 'pretty much the same thing' was the very theme of this article.

'I didn't understand a single line of the settings I left entirely to the AI.'

This is my real sin this time. No, perhaps it was an incident that happened precisely because I had Bypass Permission Mode turned on.

Chapter 5: The Strategy of Proof of Humanity 5:5

"Always check the official documentation before writing configuration files." (I won't lecture you like that.)

That's not what I'm trying to say.

There are people out there doing it with an 'AI 8: Self 2' ratio.

They use AI as a tool while injecting 20% of their own experience, judgment, and intuition.

My ratio is 'Self 5: AI 5'.

This is a strategic choice.

Whether in a professional kitchen or in the field of AI, this principle remains the same.

There is a saying I heard from a senior chef.

"The number of times you've filleted a fish with a knife becomes your persuasiveness. You can only speak about what you've done with your own hands."

"Show them"—this only holds true once you have experienced it yourself.

The more you increase the ratio of tasks left to AI, the thinner the density of your experience becomes.

When the density of experience thins, the 'heat' that reaches the reader disappears.

You end up with 'cold takoyaki'.

The moment I understood the meaning of `disable-model-invocation: true`, fixed it, and the agent team started up normally—I had that experience.

I failed, I noticed, and I fixed it.

This series of experiences is the reason I can write this article today.

AI can do research. AI can create structures.

But the primary information that 'I actually did it' is something AI can never replace.

That is human proof.

Failure experiences are the most powerful fuel for AI.

The reason I can talk about this failure is because I failed.

That alone becomes the ultimate way to survive.

The moment I confirmed this, something clicked for me.

'Knowing' is different from 'being able to use'.

From 'being able to use' to 'understanding'—only those who move in this direction can truly master AI.

Conclusion/CTA

Review one line of your prompt's activation trigger right now.

That's not all.

Look up the meaning of that setting, just one line.

'I wrote the settings' and 'I understood the settings' are two different things.

Open the configuration file you wrote and try explaining each line out loud.

If there is a line you cannot explain—that is the seed of your next failure.For me, it was almost all of them.

Within today, check it using official documentation or AI, and understand the meaning of just one line.

That is all it takes. For us non-engineers, that is a big step forward.

'Come report to me when you've done it. I'll praise you.'

Isoroku Yamamoto also said, 'If you do not praise them, people will not move.'

I will put it into practice.

Yee-haw!!

Failure experiences are the best way to survive.

It's okay to be covered in mud. It's okay to look pathetic.

The primary information you experience yourself—that is human proof.

#AI #GenerativeAI #ClaudeCode #Obsidian #PromptEngineering #FailureStory #40s #WorkEfficiency #HumanProof #API #GoroMajima #AutomationTrap

*The persona and expressions in this content are a unique parody generated by AI, inspired by a specific copyrighted work, and have no relation to any actual rights holders or organizations.