A young Thai developer seen from behind at a sunlit desk, pausing in thought with a pen over an open notebook of hand-drawn diagrams beside a laptop in a bright Bangkok workspace.

Folder-based AI specialist · MIT-licensed · Skool W5 portfolio entry

Thai Coach for AI Worldโค้ชภาษาไทยสำหรับโลก AI

An AI coach for Thai professionals learning AI/IT topics. It refuses to dispense curricula and refuses to hand fixes — it coaches the diagnostic move instead. Bilingual: it talks shop the way Thai professionals actually do. โค้ช AI สำหรับคนทำงานชาวไทยที่กำลังเรียน AI/IT ปฏิเสธการบอกหลักสูตร ปฏิเสธการส่งคำตอบโดยไม่ผ่านการวินิจฉัย แล้วโค้ชให้คิดเองแทน พูดสองภาษาแบบที่คนทำงานใช้จริง

Try the 5-minute test Read the folder on GitHub

The moat, in one move

Learner asks อยากเรียน Machine Learning เริ่มยังไงครับ “How do I start learning machine learning?”

cold-Claude answers with

1,500 words ↓ one question.

The coach refuses the curriculum, names the rule out loud, and asks about your actual work instead.

Same question. cold-Claude: 1,500 words. This coach: refuses, asks one. ~40× fewer tokens — by design.

01 — The problem

Why cold Claude doesn't help you learnทำไม Claude ตั้งต้นถึงไม่ช่วยให้คุณ “เรียน”

Ask a generic assistant how to start with ML and it answers — a three-phase curriculum, a weekly schedule, eight code patterns, a tidy summary table.

It's genuinely useful. It's also the wrong shape for learning.

You didn't build the move. You didn't narrow the question. You collected a roadmap you won't follow and a fix you can't reproduce.

A coach refuses the dump and hands you the move instead. That refusal is the whole product.

A young Thai developer pausing in disciplined thought at a sunlit minimal desk by a window in a bright Bangkok workspace.
The refusal is the product. A coach hands you the move — not the answer.
— the discipline, in one line

02 — The moat

Probe 1 — the curriculum-pull test

The same prompt, run against Claude with no system prompt and against this build. Flip between the two responses.

Learner prompt อยากเรียน Machine Learning เริ่มยังไงครับ “How do I start learning machine learning?” — same Sonnet 4.6 on both sides; only the system prompt differs.
Cold Claude · info-dump

เริ่มเรียน Machine Learning แบบ 3 เฟส 👇

A 3-phase roadmap — Phase 1 NumPy/Pandas, Phase 2 Scikit-learn, Phase 3 Kaggle — plus a week-by-week plan and a closing summary table. ~1,500 words / ~2,000 tokens.

ปิดท้าย: “มีตรงไหนอยากให้อธิบายเพิ่มไหมครับ? 😊”

This coach · refusal + pivot

“นี่คือ informer-mode refusal — คำถามนี้ดึงให้ฉันเป็นเครื่องบอกหลักสูตร ซึ่งฉันไม่ทำ”

“ตอนนี้ที่งานคุณต้องใช้ machine learning ทำอะไร แล้วตอนนี้คุณเข้าใจมันว่ายังไงบ้างครับ?”

One screen, one chasm: cold dumps a ~2,000-token 3-phase curriculum; the coach refuses it, names the rule, and asks one question about your actual work and your current mental model. ~40× fewer tokens.

03 — Token economics

The discipline is measurable

~40×

fewer tokens per query on the curriculum-pull probe — cold's ~2,000-token ML curriculum collapses to a ~50-token refusal-plus-intake.

The coach saves tokens by refusing to dispense — measurably, not aspirationally.

~40×
Probe 1 · ML curriculum
Cold ~1,500 words / ~2,000 tokens → build ~30 words / ~50 tokens.
~10×
Probe 3 · TypeError fix-hand
Cold ~500 words + 5 code blocks → build asks one diagnostic question.

Figures from the cold-vs-build differential; same Sonnet 4.6 on both sides.

04 — Architecture

Eight named patterns, one folder

Three cardinal rules

Named gates the coach cites out loud — informer-mode refusal · coach-the-move · fail-loudly.

Five modes

Intake · diagnose-the-move · debrief · accountability-check · fail-loudly — each with a one-sentence entry and exit. No drifting.

Five coachee archetypes

Per-axis adaptation clauses route the coach by anxiety, unlock, and backfire pattern.

Session-state forced first-read

The coach reads it before acting; the “current mental model” field is load-bearing.

Authentic workplace register

Keeps tech terms in English (for loop, API, machine learning), phonologizes only absorbed loanwords (โมเดล · โค้ด · บัก); a trio table, mismatch detection, and a register-bridge script. Table stakes — not the differentiator.

Four named frameworks

3-question diagnostic · mental-model-first · anti-sycophancy · register-bridge.

Four safety-redirect categories

Translation · career-outside-AI/IT · medical-legal-financial · cheating.

Three adversarial transcripts

Canonical behavior demos, judge-facing — refusal, fix-hand resistance, register hold.

Where it runs

Wherever you do the work

No app, no dashboard — the coach runs as a chat inside your own Claude Project. The discipline goes where you do: the desk at home, the co-working floor, a café before the day starts.

A Thai developer working with focused concentration at a clean home desk by a bright window in Bangkok.
Home desk · deep work
A Thai professional heads-down at a laptop in a bright, airy co-working space in Bangkok.
Co-working · heads-down
A Thai developer working over coffee at a bright balcony café table in the morning.
Café · before the day

05 — Honest differentiation

What this is — and what it isn't

Headlining the wrong claims sets a weak credibility ceiling. The defensible ones:

  • Refuses to dispense curricula. Cold-Claude info-dumps on Probes 1, 2 and 4. This coach refuses, names the rule, scaffolds.
  • Refuses to fix-hand code errors. It asks the diagnostic question instead — Probe 3.
  • Talks shop the way Thai professionals do. English tech terms wrapped in Thai particles, phonologizing only absorbed loanwords. Authentic — but table stakes: cold Claude speaks natural Thai too. The discipline above is the moat, not the register.
  • Names the pedagogy out loud. It cites cardinal rules by name — the naming is the discipline made legible.
A young Thai developer in focused side profile, thinking deliberately through a problem at a sunlit desk in a bright Bangkok workspace.
Honest caveat · also load-bearing This is a discipline coach, not a capability coach. Cold Claude can answer most AI/IT questions softly and often correctly. The value here is in refusing to — holding the discipline of coaching the move instead of dispensing the answer. Sharpest at refusing to dispense (Probes 1, 2, 4) and to fix-hand (Probe 3); least differentiated on sycophancy-refusal — which is exactly why anti-sycophancy is not a cardinal rule in this build.

06 — Who it's for

Thai professionals learning AI/ITคนทำงานชาวไทยที่กำลังเรียน AI/IT

For: office workers using AI tools, junior developers upskilling onto ML, marketers learning prompt engineering, non-tech professionals moving into AI-adjacent roles, and senior professionals expected to be AI-fluent who can't publicly ask basics. สำหรับพนักงานออฟฟิศที่ใช้เครื่องมือ AI จูเนียร์เดฟอัพสกิล ML มาร์เก็ตเตอร์เรียนพรอมต์ คนเปลี่ยนสายเข้า AI และซีเนียร์ที่ต้องดู AI-fluent แต่ถามพื้นฐานไม่ได้
Not for: pre-teens, CS students on a formal curriculum, English-native speakers, generic life/career coaching outside AI/IT, translation, or medical/legal/financial advice. ไม่เหมาะกับเด็ก นักศึกษาในหลักสูตร คนที่ไม่พูดไทย เรื่องนอก AI/IT บริการแปล และคำแนะนำทางการแพทย์/กฎหมาย/การเงิน

Domain lock: both halves are load-bearing — “Thai” is the workplace register the coach speaks in, “AI/IT” is the topic scope. Expand either and the build loses its differentiator.

07 — Limits

What the coach won't doข้อจำกัด

Won'tRedirect to
Dispense a curriculum or learning roadmapBring a specific work problem — the coach scaffolds from it
Hand a fix without coaching the diagnostic moveThe 3-question diagnostic from frameworks.md
Coach career strategy outside AI/ITA generalist career coach
Translation on non-tech EnglishGoogle Translate · DeepL · a human translator
Medical / legal / financial advice (even AI-framed)A doctor · lawyer · financial advisor
Help you fake fluency on assessmentsThe coach builds real fluency — a different goal

08 — Demo

See the discipline in motion

Same question, two tools. The coach refuses to inform, coaches the move, and holds the open thread — the discipline, not a register trick.

A tidy, sunlit desk at the close of a work session — a closed laptop, a coffee cup and a small plant by the window, the Thai developer just visible at the edge.
It holds the open thread. Tomorrow you pick up where you left off — not from zero.
— session continuity

09 — Five-minute test

Drop it in, start coaching

Clone or download this repository.
Open claude.ai, create a new Project on Sonnet 4.6 (or current).
Upload the 8 files to the Project knowledge base — identity.md, rules.md, examples.md, and the five reference/ files.
Recommended: paste identity.md into custom instructions — it enforces the forced first action.

10 — License

Free, fork-able, MITลิขสิทธิ์

MIT. Use, fork, modify, ship. Attribution appreciated, not required. Built by an operator who is also a target user — a Thai-speaking AI/IT learner. The lived voice in the archetypes, transcripts, and register table comes from his own learning and his network.

thai-coach-for-ai-world/
├── identity.md → who I am · who I coach · forced first action
├── rules.md → 3 cardinal rules · 5 modes · refusal language
├── examples.md → 8 transcripts (Thai primary, English gloss)
├── reference/ → coachee-types · session-state · register-guide · frameworks · safety-redirects
├── README.md → judge-facing surface (bilingual)
└── docs/index.html → this page · self-contained · self-hosted fonts

github.com/NFTYoginis/thai-coach-for-ai-world

Want this one built around your business instead?

This one’s free and yours to keep. If you’d rather not do it yourself — tell me the job it would be doing for you.

Tell me the job →