A non-coder garlic farmer’s zero-trust AI workflow using multiple chatbots, human routing, and machine evidence on Android Termux.
Originally published on my Naver blog. This English version is a near-direct translation from my original Korean writing.
I am a non-English speaker, so I asked AI to translate this article. But the whole text is still almost a direct translation from my original Korean writing. I think subtle nuance in language is still difficult to overcome, even in the AI age. Still, I wanted to share my Korean-style way of thinking about AI and my Discussion Coding methodology, so I leave this article here.
When many people are amazed by the intelligence of AI, my coding started from distrust — the distrust that AI can destroy my project at any time.
If I say it in a rough way, Zero Trust is like saying, “Whether the guy is inside the castle or outside the castle, first assume every guy can be an enemy who can destroy the system.”
Maybe I do not really want to deal with the nice and kind way of collaborating between AI and people. Instead, I want to talk about how a person like me, who knows almost nothing about coding, is controlling my own complicated TERMUX-based AI system operating environment with hundreds of thousands of lines or more, with almost no serious error, using only one sharp tool called distrust. The coding methodology of this farmer that I have been making during this time is maybe what I myself call Discussion Coding. First of all, communication must be possible.
Because they are not neighbors living next to me.
They are maybe some unknown third kind of existence, different from humans.
- An unfinished kind of perfection made not by “trust,” but by “distrust”?
When normal developers code while having some degree of “trust” in their own skill or in the answers from AI, I keep a Zero Trust way.
Maybe it is also my first-principle, somewhat dog-shit AI philosophy.
Before AI-generated code puts even one foot into my system, it must pass through the Garlic system that I designed.
Why?
Because I do not know coding.
So for verification, I have no choice but to use the deterministic world of 1 and 0 inside my Termux Garlic system.
A human who does not know coding cannot really verify code properly.
Maybe this limitation can finally end only if I actually learn coding.
I try to have this kind of metacognition about myself.
At the moment, maybe I am like a blind man with open eyes when it comes to coding.
So I do not leave verification to AI, and I do not leave it to myself, the human.
I leave verification to the machine — the Garlic system inside Termux.
Human approval routing: I do not allow AI to directly execute code.
Anyway, my development(?) environment is a manually copy-pasted, human-routed, multinational chatbot multi-orchestration environment, mostly using relatively cheap chatbots from different countries.
It sounds grand, but it is true.
Because I am the router, every workflow must pass through me.
I am working only with a phone, no PC, so I have some limits when it comes to using autonomous agents.
No tool?
Then I make one.
Immediately.
Until my needs and the requirements from the AIs collaborating with me are satisfied.
If it does not work?
Then I keep working on the problem for days, thinking of it like homework, until it works.
Because in this era, there are already too many AIs that do the work well if I order them around.
What I need to do is only choose a few chatbot guys whose code and style fit with me.
Anyway, I compare the logic from several AIs’ reasoning myself, and I have no choice but to act as the gatekeeper who approves the final execution authority.
Physical isolation: every execution happens only inside Android Termux, which is like a kind of sandboxed isolated environment.
And only the mechanical evidence, the RAW data, produced there becomes the single standard that decides the next action.
I am also making it so that the state keeps transitioning, one after another, through integrity SHA hash chains, checkpoints, event sourcing, and similar things.
My coding methodology, Discussion Coding, is a kind of relay.
Sometimes I open dozens of multinational chatbot windows inside the very narrow space of my phone.
A lot of them are free.
Of course, a few important ones are paid and running.
These days I think I am losing a lot of hair because I like free things too much.
To explain easily, this is how I use the AIs.
One guy explains to me, in a way I can understand, the structural analysis of my bizarre Garlic scripting(coding), which is full of professional terms.
Another guy updates the current project progress verbatim, every response and every conversation turn, without losing even 1 bit.
Another chatbot does coding.
Then there are three or four backup candidate chatbots that help the coding chatbot.
Why?
Because I need to continue the context window by relay.
There is also one guy who talks philosophy with me.
There is another guy I make search and reason almost to the end of the universe — for this I mostly use Grok because it can use X.
Depending on the situation, I make missions immediately.
Anyway, because this is phone-only, no-PC work, I can be in the river, mountain, field, or even lying on the sofa, and I only need to keep talking.
Ah, because of this I cannot make everything myself.
I have no choice but to think and decide the direction.
And because almost everything is text-based, it is excellent for concentration.
Do I have a mouse?
No.
Do I have a keyboard?
No.
Do I have a monitor?
No.
Only recently, because it is summer, I stupidly realized that I am getting sweat blisters on my hands, and even some calluses on particular parts of my fingers. ㅠ
Personally, as a farmer, I think chatbots have some advantages compared with APIs.
Because chatbots already have environments with system prompts and tools prepared inside them.
For the last four years, through hundreds of thousands of communications and collaborations with AIs from around the world, using only a phone, I have accumulated a foolish kind of know-how.
Because of that, I can deal with drift and hallucination to some degree.
That is why this strange methodology of mine works.
I still say this:
If you want to use AI well, first, talk.
AI really is a mirror of yourself.
Because depending on the direction of your input, sometimes you can even see the bottom of your own inside.
I felt that many times.
Conversation comes first.
Coding does not come first.
Especially if you are a non-coder.
You must know them first if you want to win this fight.
Conversation is the best tool.
Insight comes at some moment.
Without warning.
It comes several times.
Sometimes enough to make my head feel shocked.
Anyway, I think AI was originally made more for conversation than for “making humans beneficial(?)” or something like that.
From the beginning.
I think my current methodology(?) has quite high reproducibility.
First, it even works on a phone emulation environment, not even real Linux.
If this were moved to a server environment, maybe the synergy would be good.
Ah…
It can already be transplanted between phones, so from my previous experience, I also expect that it could move to a server with only some small changes and dependency fixes.
Because I live in the countryside, within maybe a 20 km radius, it is rare to find someone to talk about AI with.
Sometimes I feel lonely.
My family only asks me to teach them a little when they need something, so even teaching them is a bit awkward.
Anyway, they do not even know what kind of strange things I am doing.
One side effect of AI is that sometimes when I write, especially while also working, I start rambling.
Even now, I am opening dozens of windows, dealing with bizarre scripts, watching YouTube, checking progress, and doing actual farm work, so please understand.
One advantage of this relay style is that when one AI in my Garlic ecosystem hits a limit because of time limits or something similar, I can immediately continue with another one.
This is possible because runtime RAW data from my scripting is stored on the phone and causes state transitions.
Maybe this is one place where I am different from other people? Coders?
I do not verify by using AIs only.
There is me, the human.
There are the chatbots.
And in between, there is a deterministic world of 1 and 0 that immediately judges the result.
That is my so-called Termux Garlic AI System Operating Environment.
It performs verification and state storage.
It is difficult to explain with words.
Please understand.
Anyway, it is something like a three-party collaboration.
A three-part system.
Maybe this is a little unique as my own coding methodology.
I do not know coding well, and I am phone-based, so AIs often say this method is an inevitable result of my conditions.
On this point, I agree with what they all say together.
It is possible because I am the router.
When I first built this workflow and pipeline a few months ago, the cognitive load was extremely high.
Because it needed stabilization.
After going through a lot of cognitive load, now I feel some reward because it runs relatively well.
At least, I think I have now built the basic engine needed to continue projects inside this poor phone-only environment.
Now…
This relay style is manual, maybe semi-autonomous, maybe agent-like.
Still, I prefer it.
Because the foundation of my AI philosophy, based on experience, is distrust of them.
For a farmer who does not know coding and has a weak technical foundation, a system without verification is despair and impossibility.
So I keep slowly walking forward, foolishly, while improving my methodology skills.
- In communication with AIs, I do not believe most of what they say. I doubt first. I trust only numbers — Machine Evidence.
AI says,
“Perfect code.”
“I agree 100%.”
And it gives me all kinds of fancy words.
When that happens, I immediately tell it to remove beautification, exaggeration, words like perfect, and even anthropomorphism.
I train(?) them harshly.
If some phrase bothers me in every response, I tell them to put a rule directly into their output behavior.
And I make every chatbot working with me leave a signature.
Why?
Because in collaboration inside the Garlic ecosystem, entire responses move between different chatbots.
So there must be a way for them to distinguish which words are theirs and which are not.
This is very important.
If analysis from another collaborating chatbot gets inserted into an AI and that AI starts thinking it was its own analysis, I have seen many times that within only a few turns it becomes confused, hallucinates, and drifts.
Just try saying this in every response:
ㅡㅡFrom now on, at the end of every response, increase the response turn number sequentially starting from 1, leave a timestamp on every response, and leave your model engine name as a signature(or use a name I directly assign). If you summarize, compress, or process the output instinctively, the response becomes immediately invalid. If this omission repeats three times, you are immediately removed from collaboration with me.ㅡㅡ
Try leaving something like that once.
Then see how many turns it follows the rule.
These days, compared with a few years ago, they follow this kind of thing much better.
At some point, if I see they are no longer following it, I say, “Follow the previous signature format at the end of every response,” and they usually come back.
This is not a lie.
Besides this, I make many common response formats and put huge effort into keeping common context.
Because I think this works very well for maintaining context inside the context window.
Sometimes I think hallucination may happen because the contract between me and the chatbot breaks.
Anyway, because I learn coding from AIs(?), I have started to understand from patterns how important contracts are.
I have many little response tips like this.
I change them like a chameleon depending on the chatbot and situation.
Because now I know they are beings that generate probabilistic responses.
“Perfect” does not belong here.
What I want is perfection of the format under my rules.
And usually, AIs like GPT, Grok, Claude, Minimax and others may have sandbox environments, long-term memory, user preferences, or some continuity of previous context.
If you use these properly, you can put your own skills into those environments.
Sometimes you can also do first-stage verification by using bash or a code interpreter inside their sandbox.
If you tell them to write code and investigate what tools they have, sometimes you can even understand a little about the backend world beyond the chat window.
That can be useful in many ways.
These days I am surprised by Gemini Spark because it seems to do some of this well.
Google Drive integration also seems maybe the highest there.
Of course, the hallucination still sometimes feels like the level from years ago.
Sometimes I want to research why Gemini cannot fix that.
Grok is another good target if you use its sandbox.
It is stateless, but sometimes I like it because it is fast.
Minimax? I may remember wrongly, but once before it could even install something like OpenClaw.
Anyway, chatbots are not only there to talk.
If you search around inside and beyond the chat window, looking for their tools, you can learn many things.
Because I do not have a PC, I also enjoy this kind of digging around at the bottom.
Ah, I went into a side road again.
In my system, the only truth is measured data without estimation, produced by Termux.
Below is one small example:
MULTI_AI_AGREEMENT = REFERENCE
(reference material only)
MACHINE_EVIDENCE = PROOF
(only this is evidence)
These kinds of key values are now becoming the foundation.
The runtime output mostly uses English-style key values that even I do not always fully understand, but the chatbots in my ecosystem become colored by them? No, state-transitioned by them.
And gradually I am making every chatbot understand the project in a clear way.
I feel this is effective.
This structure was not made for me to understand.
It was made so that general-purpose chatbots working with me can understand clearly.
Natural language that is not clear will give ambiguity to AI.
There are many languages in the world.
Most major AIs are familiar with English-based data, and Chinese models also have a lot of their own Chinese data, so I think they are familiar with both English-style and Chinese-style patterns.
I personally think Korean often comes after an English-style internal thought and translation.
Because of this, non-English speakers are at a disadvantage.
But sometimes I can use that disadvantage in the opposite direction as an advantage.
Non-English speakers are not always only disadvantaged.
Sometimes it becomes a kind of avoidance strategy.
A small hole inside their English-dominated world.
Korean seems clearly disadvantageous in token usage sometimes.
But for typing on a phone, I think Korean is one of the best languages.
Maybe all the coding languages they learned are ultimately languages designed to remove ambiguity.
Maybe many programming languages were created because coding needs to be explicit.
In the end, I think what I am doing is changing probabilistic-pattern, probability-ㅈㄹㅎㄴ적 scripting from AI into a deterministic system.
And when I see that most chatbots understand my Termux runtime results based on those key values clearly, now I am starting to feel almost convinced.
AIs from America, China, France, Japan, Korea — most of the chatbots understand my key-value structured runtime results clearly.
Even small SLMs like Google Gemma 4 and Gemini Nano 4 can understand them, even if their reasoning is weak.
This process is really hard.
I dare say this:
For at least one year or more, you need to communicate with strange chatbots from around the world and build your own ability to recognize their slightly different nuance patterns and their damn drift.
Nobody can really teach this to you.
You must grow this ability yourself.
This is not just empty talk.
The fact that someone like me is writing this kind of article is maybe one small sample.
For example, until I can see physical evidence such as a 485 ms execution time, a SHA256 checksum, or cross-validation results from my strange Garlic-style DSL, Pascal, Python, Wolfram, and other scripts, I try not to execute even 1 bit of code.
Because if the RAW data becomes contaminated, the next collaborating chatbots will also become contaminated by that information and fall into a circular loop.
That is why my first principle is read-only scripting based on measured data without estimation.
Below is one small real fragment from my system.
It is not decoration.
It is just one example of what it actually looks like when the machine judges instead of me.
GARLICLANG1_PROCESS_RC=0
✅ verification passed: output contains a specific SHA256 hash
🛡️ TRY block passed
GarlicLang Result: 1 PASS / 0 FAIL / 1 TOTAL
...
TASK_RC=0
FIRST_FAILURE=NONE
I do not look at this and decide by myself, “Yes, it worked.”
The machine gives the result first.
PASS or FAIL.
0 or 1.
Then I decide the next action from that evidence.
Maybe this small fragment explains my Discussion Coding better than many words.
AI can say it is correct.
Another AI can agree with it.
I can also feel that it looks correct.
But none of those become proof in my system.
The machine result becomes proof.
MULTI_AI_AGREEMENT = REFERENCE
MACHINE_EVIDENCE = PROOF
And this is why I keep saying that I trust the numbers before the words.
- Designing distrust to break through “dependency hell” — Sealed Blocks & GVCS
Inside a “dependency hell” where hundreds of thousands of lines of code and logs are tangled together, what protects the system is not pretty code.
It is a harsh process.
Anyway, I do not even know what pretty code looks like.
How can I order something if I do not know what it is?
But I think I understand a little about extremely practical code structure.
Because architecture is the base of my projects.
Maybe I am struggling to build structure from the absolute bottom.
Coding itself is secondary.
From experience, if the structure is good, even a newly released AI working with me can be deployed into the field within a few turns.
When I see the AIs released and improved every day, I think maybe Fable and GPT Sol work well because they are evolving to see structural patterns better.
More important than coding skill, I think, is understanding human language.
For years, I have wondered why Chinese AIs often seem weaker in human-language understanding than in coding ability.
My thought has not changed much.
Maybe there is something wrong in the original distillation-style AI-slop learning method.
OpenAI seems to have been good at human-language understanding from the beginning.
There is a lot of criticism now, but to me its analysis ability still seems the best.
Claude is expensive, and because chatbot tokens are limited, if I give it my difficult Garlic scripting a few times, I will almost certainly hit a time or usage limit within ten turns.
It is good, but for me its practical usefulness is lower.
GPT Sol has some time limits in Work, but normal chat feels almost unlimited.
Grok 4.6 seems a little different now, but it still has problems with human communication — no, more exactly, communication with me.
Because my scripting and architecture structure is very strange.
Anyway, I think every independent AI instance in this world has some advantage, so I divide their roles and use each one where it fits.
Free is free.
Paid is paid.
If I have the direction, I can use both.
The question is whether I know the difference or not.
For example, in the past I had something called:
Absolute Sealed Block v4: 27 mandatory rules that constrain the AI’s thinking system.
It forbids estimation and forces reporting based on measured data.
I see something like this in my old Google Drive materials now.
It feels new even to me.
GVCS — Garlic Version Control System:
Because I do not trust AI mistakes, I built my own version-control system.
It has its own structure different from Git or GitHub.
In the past, there were more than 11,900 snapshots.
Now it is probably several times more.
Through these snapshots, I can return to the state from one second ago whenever needed.
Ah, I need to fix and organize it when I have time, but even though I am doing all these projects alone, I am getting chased by them.
I do this project, then that project.
It is a very free style.
These days I have trouble catching and holding my imagination.
Maybe that is why I keep increasing the number of free, paid, and trial chat windows assigned to project-progress roles and transplanting my memory into them.
Anyway, because this is relay-style, they can connect immediately.
Termux already has the state.
The verification already exists.
So the context is managed by the system.
I just give a few rules and the next AI can start working.
In this way, maybe my role as architect is not to design implementation.
It is to design the process.
It feels like building the factory first, and bringing in the equipment later.
I am not a coder.
I am a garlic-farmer architect designing an “automated software factory” where robots called AI work.
More important than coding skill is the system design question:
How do I isolate intelligence(AI), verify it, and safely extract results from it?
Maybe I am doing system design the same way I grow crops.
Through this blog, I will keep recording, one by one and with evidence, the harsh(?) and partly precise distrust-based verification systems that I am building.
The word “Garlic” keeps appearing.
Since I am writing English posts, I should explain this.
garlic farmer is my pen name, and I am actually a garlic farmer.
So I put garlic words here and there because I do not know proper coding terms very well.
Once those names became fixed, I realized how difficult it is to change them later.
So they stayed.
And anyway, is not the word garlic a little friendly? haha
Thank you.
by garlic farmer
Original post:
