{"slug": "no-vibe-coding-discussion-coding-on-termux-phone-only-zero-trust", "title": "No Vibe Coding: Discussion Coding on Termux – Phone Only, Zero Trust", "summary": "A non-English-speaking farmer-developer describes a 'Discussion Coding' methodology for managing a Termux-based AI system on a phone, built on Zero Trust principles and manual copy-paste routing of multiple chatbots. The approach relies on a custom 'Garlic' verification system and physical isolation in Android Termux, with human approval for all code execution.", "body_md": "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.\n\nWhen 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.\n\nIf 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.”\n\nMaybe 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.\n\nBecause they are not neighbors living next to me.\n\nThey are maybe some unknown third kind of existence, different from humans.\n\n1. An unfinished kind of perfection made not by “trust,” but by “distrust”?\n\nWhen 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.\n\nMaybe it is also my first-principle, somewhat dog-shit AI philosophy.\n\nBefore AI-generated code puts even one foot into my system, it must pass through the Garlic system that I designed.\n\nWhy?\n\nBecause I do not know coding.\n\nSo for verification, I have no choice but to use the deterministic world of 1 and 0 inside my Termux Garlic system.\n\nA human who does not know coding cannot really verify code properly.\n\nMaybe this limitation can finally end only if I actually learn coding.\n\nI try to have this kind of metacognition about myself.\n\nAt the moment, maybe I am like a blind man with open eyes when it comes to coding.\n\nSo I do not leave verification to AI, and I do not leave it to myself, the human.\n\nI leave verification to the machine — the Garlic system inside Termux.\n\nHuman approval routing: I do not allow AI to directly execute code.\n\nAnyway, my development(?) environment is a manually copy-pasted, human-routed, multinational chatbot multi-orchestration environment, mostly using relatively cheap chatbots from different countries.\n\nIt sounds grand, but it is true.\n\nBecause I am the router, every workflow must pass through me.\n\nI am working only with a phone, no PC, so I have some limits when it comes to using autonomous agents.\n\nNo tool?\n\nThen I make one.\n\nImmediately.\n\nUntil my needs and the requirements from the AIs collaborating with me are satisfied.\n\nIf it does not work?\n\nThen I keep working on the problem for days, thinking of it like homework, until it works.\n\nBecause in this era, there are already too many AIs that do the work well if I order them around.\n\nWhat I need to do is only choose a few chatbot guys whose code and style fit with me.\n\nAnyway, 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.\n\nPhysical isolation: every execution happens only inside Android Termux, which is like a kind of sandboxed isolated environment.\n\nAnd only the mechanical evidence, the RAW data, produced there becomes the single standard that decides the next action.\n\nI am also making it so that the state keeps transitioning, one after another, through integrity SHA hash chains, checkpoints, event sourcing, and similar things.\n\nMy coding methodology, Discussion Coding, is a kind of relay.\n\nSometimes I open dozens of multinational chatbot windows inside the very narrow space of my phone.\n\nA lot of them are free.\n\nOf course, a few important ones are paid and running.\n\nThese days I think I am losing a lot of hair because I like free things too much.\n\nTo explain easily, this is how I use the AIs.\n\nOne 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.\n\nAnother guy updates the current project progress verbatim, every response and every conversation turn, without losing even 1 bit.\n\nAnother chatbot does coding.\n\nThen there are three or four backup candidate chatbots that help the coding chatbot.\n\nWhy?\n\nBecause I need to continue the context window by relay.\n\nThere is also one guy who talks philosophy with me.\n\nThere 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.\n\nDepending on the situation, I make missions immediately.\n\nAnyway, 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.\n\nAh, because of this I cannot make everything myself.\n\nI have no choice but to think and decide the direction.\n\nAnd because almost everything is text-based, it is excellent for concentration.\n\nDo I have a mouse?\n\nNo.\n\nDo I have a keyboard?\n\nNo.\n\nDo I have a monitor?\n\nNo.\n\nOnly 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\n\nPersonally, as a farmer, I think chatbots have some advantages compared with APIs.\n\nBecause chatbots already have environments with system prompts and tools prepared inside them.\n\nFor 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.\n\nBecause of that, I can deal with drift and hallucination to some degree.\n\nThat is why this strange methodology of mine works.\n\nI still say this:\n\nIf you want to use AI well, first, talk.\n\nAI really is a mirror of yourself.\n\nBecause depending on the direction of your input, sometimes you can even see the bottom of your own inside.\n\nI felt that many times.\n\nConversation comes first.\n\nCoding does not come first.\n\nEspecially if you are a non-coder.\n\nYou must know them first if you want to win this fight.\n\nConversation is the best tool.\n\nInsight comes at some moment.\n\nWithout warning.\n\nIt comes several times.\n\nSometimes enough to make my head feel shocked.\n\nAnyway, I think AI was originally made more for conversation than for “making humans beneficial(?)” or something like that.\n\nFrom the beginning.\n\nI think my current methodology(?) has quite high reproducibility.\n\nFirst, it even works on a phone emulation environment, not even real Linux.\n\nIf this were moved to a server environment, maybe the synergy would be good.\n\nAh…\n\nIt 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.\n\nBecause I live in the countryside, within maybe a 20 km radius, it is rare to find someone to talk about AI with.\n\nSometimes I feel lonely.\n\nMy family only asks me to teach them a little when they need something, so even teaching them is a bit awkward.\n\nAnyway, they do not even know what kind of strange things I am doing.\n\nOne side effect of AI is that sometimes when I write, especially while also working, I start rambling.\n\nEven now, I am opening dozens of windows, dealing with bizarre scripts, watching YouTube, checking progress, and doing actual farm work, so please understand.\n\nOne 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.\n\nThis is possible because runtime RAW data from my scripting is stored on the phone and causes state transitions.\n\nMaybe this is one place where I am different from other people? Coders?\n\nI do not verify by using AIs only.\n\nThere is me, the human.\n\nThere are the chatbots.\n\nAnd in between, there is a deterministic world of 1 and 0 that immediately judges the result.\n\nThat is my so-called Termux Garlic AI System Operating Environment.\n\nIt performs verification and state storage.\n\nIt is difficult to explain with words.\n\nPlease understand.\n\nAnyway, it is something like a three-party collaboration.\n\nA three-part system.\n\nMaybe this is a little unique as my own coding methodology.\n\nI do not know coding well, and I am phone-based, so AIs often say this method is an inevitable result of my conditions.\n\nOn this point, I agree with what they all say together.\n\nIt is possible because I am the router.\n\nWhen I first built this workflow and pipeline a few months ago, the cognitive load was extremely high.\n\nBecause it needed stabilization.\n\nAfter going through a lot of cognitive load, now I feel some reward because it runs relatively well.\n\nAt least, I think I have now built the basic engine needed to continue projects inside this poor phone-only environment.\n\nNow…\n\nThis relay style is manual, maybe semi-autonomous, maybe agent-like.\n\nStill, I prefer it.\n\nBecause the foundation of my AI philosophy, based on experience, is distrust of them.\n\nFor a farmer who does not know coding and has a weak technical foundation, a system without verification is despair and impossibility.\n\nSo I keep slowly walking forward, foolishly, while improving my methodology skills.\n\n2. In communication with AIs, I do not believe most of what they say. I doubt first. I trust only numbers — Machine Evidence.\n\nAI says,\n\n“Perfect code.”\n\n“I agree 100%.”\n\nAnd it gives me all kinds of fancy words.\n\nWhen that happens, I immediately tell it to remove beautification, exaggeration, words like perfect, and even anthropomorphism.\n\nI train(?) them harshly.\n\nIf some phrase bothers me in every response, I tell them to put a rule directly into their output behavior.\n\nAnd I make every chatbot working with me leave a signature.\n\nWhy?\n\nBecause in collaboration inside the Garlic ecosystem, entire responses move between different chatbots.\n\nSo there must be a way for them to distinguish which words are theirs and which are not.\n\nThis is very important.\n\nIf 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.\n\nJust try saying this in every response:\n\nㅡㅡFrom 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\n\nTry leaving something like that once.\n\nThen see how many turns it follows the rule.\n\nThese days, compared with a few years ago, they follow this kind of thing much better.\n\nAt 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.\n\nThis is not a lie.\n\nBesides this, I make many common response formats and put huge effort into keeping common context.\n\nBecause I think this works very well for maintaining context inside the context window.\n\nSometimes I think hallucination may happen because the contract between me and the chatbot breaks.\n\nAnyway, because I learn coding from AIs(?), I have started to understand from patterns how important contracts are.\n\nI have many little response tips like this.\n\nI change them like a chameleon depending on the chatbot and situation.\n\nBecause now I know they are beings that generate probabilistic responses.\n\n“Perfect” does not belong here.\n\nWhat I want is perfection of the format under my rules.\n\nAnd usually, AIs like GPT, Grok, Claude, Minimax and others may have sandbox environments, long-term memory, user preferences, or some continuity of previous context.\n\nIf you use these properly, you can put your own skills into those environments.\n\nSometimes you can also do first-stage verification by using bash or a code interpreter inside their sandbox.\n\nIf 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.\n\nThat can be useful in many ways.\n\nThese days I am surprised by Gemini Spark because it seems to do some of this well.\n\nGoogle Drive integration also seems maybe the highest there.\n\nOf course, the hallucination still sometimes feels like the level from years ago.\n\nSometimes I want to research why Gemini cannot fix that.\n\nGrok is another good target if you use its sandbox.\n\nIt is stateless, but sometimes I like it because it is fast.\n\nMinimax? I may remember wrongly, but once before it could even install something like OpenClaw.\n\nAnyway, chatbots are not only there to talk.\n\nIf you search around inside and beyond the chat window, looking for their tools, you can learn many things.\n\nBecause I do not have a PC, I also enjoy this kind of digging around at the bottom.\n\nAh, I went into a side road again.\n\nIn my system, the only truth is measured data without estimation, produced by Termux.\n\nBelow is one small example:\n\nMULTI_AI_AGREEMENT = REFERENCE\n\n(reference material only)\n\nMACHINE_EVIDENCE = PROOF\n\n(only this is evidence)\n\nThese kinds of key values are now becoming the foundation.\n\nThe 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.\n\nAnd gradually I am making every chatbot understand the project in a clear way.\n\nI feel this is effective.\n\nThis structure was not made for me to understand.\n\nIt was made so that general-purpose chatbots working with me can understand clearly.\n\nNatural language that is not clear will give ambiguity to AI.\n\nThere are many languages in the world.\n\nMost 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.\n\nI personally think Korean often comes after an English-style internal thought and translation.\n\nBecause of this, non-English speakers are at a disadvantage.\n\nBut sometimes I can use that disadvantage in the opposite direction as an advantage.\n\nNon-English speakers are not always only disadvantaged.\n\nSometimes it becomes a kind of avoidance strategy.\n\nA small hole inside their English-dominated world.\n\nKorean seems clearly disadvantageous in token usage sometimes.\n\nBut for typing on a phone, I think Korean is one of the best languages.\n\nMaybe all the coding languages they learned are ultimately languages designed to remove ambiguity.\n\nMaybe many programming languages were created because coding needs to be explicit.\n\nIn the end, I think what I am doing is changing probabilistic-pattern, probability-ㅈㄹㅎㄴ적 scripting from AI into a deterministic system.\n\nAnd 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.\n\nAIs from America, China, France, Japan, Korea — most of the chatbots understand my key-value structured runtime results clearly.\n\nEven small SLMs like Google Gemma 4 and Gemini Nano 4 can understand them, even if their reasoning is weak.\n\nThis process is really hard.\n\nI dare say this:\n\nFor 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.\n\nNobody can really teach this to you.\n\nYou must grow this ability yourself.\n\nThis is not just empty talk.\n\nThe fact that someone like me is writing this kind of article is maybe one small sample.\n\nFor 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.\n\nBecause if the RAW data becomes contaminated, the next collaborating chatbots will also become contaminated by that information and fall into a circular loop.\n\nThat is why my first principle is read-only scripting based on measured data without estimation.\n\n3. Designing distrust to break through “dependency hell” — Sealed Blocks & GVCS\n\nInside a “dependency hell” where hundreds of thousands of lines of code and logs are tangled together, what protects the system is not pretty code.\n\nIt is a harsh process.\n\nAnyway, I do not even know what pretty code looks like.\n\nHow can I order something if I do not know what it is?\n\nBut I think I understand a little about extremely practical code structure.\n\nBecause architecture is the base of my projects.\n\nMaybe I am struggling to build structure from the absolute bottom.\n\nCoding itself is secondary.\n\nFrom experience, if the structure is good, even a newly released AI working with me can be deployed into the field within a few turns.\n\nWhen 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.\n\nMore important than coding skill, I think, is understanding human language.\n\nFor years, I have wondered why Chinese AIs often seem weaker in human-language understanding than in coding ability.\n\nMy thought has not changed much.\n\nMaybe there is something wrong in the original distillation-style AI-slop learning method.\n\nOpenAI seems to have been good at human-language understanding from the beginning.\n\nThere is a lot of criticism now, but to me its analysis ability still seems the best.\n\nClaude 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.\n\nIt is good, but for me its practical usefulness is lower.\n\nGPT Sol has some time limits in Work, but normal chat feels almost unlimited.\n\nGrok 4.6 seems a little different now, but it still has problems with human communication — no, more exactly, communication with me.\n\nBecause my scripting and architecture structure is very strange.\n\nAnyway, I think every independent AI instance in this world has some advantage, so I divide their roles and use each one where it fits.\n\nFree is free.\n\nPaid is paid.\n\nIf I have the direction, I can use both.\n\nThe question is whether I know the difference or not.\n\nFor example, in the past I had something called:\n\nAbsolute Sealed Block v4: 27 mandatory rules that constrain the AI’s thinking system.\n\nIt forbids estimation and forces reporting based on measured data.\n\nI see something like this in my old Google Drive materials now.\n\nIt feels new even to me.\n\nGVCS — Garlic Version Control System:\n\nBecause I do not trust AI mistakes, I built my own version-control system.\n\nIt has its own structure different from Git or GitHub.\n\nIn the past, there were more than 11,900 snapshots.\n\nNow it is probably several times more.\n\nThrough these snapshots, I can return to the state from one second ago whenever needed.\n\nAh, 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.\n\nI do this project, then that project.\n\nIt is a very free style.\n\nThese days I have trouble catching and holding my imagination.\n\nMaybe 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.\n\nAnyway, because this is relay-style, they can connect immediately.\n\nTermux already has the state.\n\nThe verification already exists.\n\nSo the context is managed by the system.\n\nI just give a few rules and the next AI can start working.\n\nIn this way, maybe my role as architect is not to design implementation.\n\nIt is to design the process.\n\nIt feels like building the factory first, and bringing in the equipment later.\n\nI am not a coder.\n\nI am a garlic-farmer architect designing an “automated software factory” where robots called AI work.\n\nMore important than coding skill is the system design question:\n\nHow do I isolate intelligence(AI), verify it, and safely extract results from it?\n\nMaybe I am doing system design the same way I grow crops.\n\nThrough 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.\n\nThe word “Garlic” keeps appearing.\n\nSince I am writing English posts, I should explain this.\n\ngarlic farmer is my pen name, and I am actually a garlic farmer.\n\nSo I put garlic words here and there because I do not know proper coding terms very well.\n\nOnce those names became fixed, I realized how difficult it is to change them later.\n\nSo they stayed.\n\nAnd anyway, is not the word garlic a little friendly? haha\n\nThank you.\n\nby garlic farmer", "url": "https://wpnews.pro/news/no-vibe-coding-discussion-coding-on-termux-phone-only-zero-trust", "canonical_source": "https://m.blog.naver.com/amadale/224399795090", "published_at": "2026-09-03 06:47:33+00:00", "updated_at": "2026-09-03 06:51:51.792823+00:00", "lang": "en", "topics": ["ai-tools", "ai-safety", "ai-agents"], "entities": ["Termux", "Android", "Garlic"], "alternates": {"html": "https://wpnews.pro/news/no-vibe-coding-discussion-coding-on-termux-phone-only-zero-trust", "markdown": "https://wpnews.pro/news/no-vibe-coding-discussion-coding-on-termux-phone-only-zero-trust.md", "text": "https://wpnews.pro/news/no-vibe-coding-discussion-coding-on-termux-phone-only-zero-trust.txt", "jsonld": "https://wpnews.pro/news/no-vibe-coding-discussion-coding-on-termux-phone-only-zero-trust.jsonld"}}