PTENES
Course · INEMA.CLUB
🧠

The AI's Hidden Manual

How to instruct models the way labs do — of the foundation to the Fable's brain.

This course distills how cutting-edge AIs— Anthropic, OpenAI, xAI, Cursor, Perplexity… — are really trained. And shows how to recover a good agent's "way of working" from the logs from your sessions—even without having the model's weights.

3
Tracks
20
Modules
120
Topics
Beg. → Adv.
Level
Layered architecture of a system prompt in neon glow — illustration from the course The Hidden Manual of AI
THE HIDDEN MANUAL 🔒 system prompt the invisible instruction that no one sees comes BEFORE everything THE 3 LAYERS OF THE COURSE 🧱 Fundamentals 🛠️ Practice 🧠 Advanced from zero jargon to the labs’ level THE BEHAVIOR 🤖 the AI acting what you SEE on the screen .jsonl logs distill the manual back

What you'll learn

Three tracks, three levels of depth. Each stands alone, but the order tells a story: first you understands the invisible note, later assembles yours, finally, does what the labs.

Who it's for

No matter where you start, there’s a path that begins at your level.

1

Beginner

Has never instructed an AI and doesn’t even know what a “prompt” is. Wants to understand, without jargon, what determines how each AI behaves. Starting from absolute zero.

Go to Track 1 →
2

Practitioner

You already use AI and want to build things for real: prompts that work, skills that trigger at the right time, disciplined agents. You want to stop improvising.

Go to Track 2 →
3

Advanced

Know the basics and want the lab-level material: read the evolution through diffs, consolidate without losing behavior, and recover an agent's method from its logs.

Go to Track 3 →
The story behind the course

We lost Fable. We recovered the method.

O Fable was a model with a rare discipline: it thought before acting, checked its own work, narrated its decisions. Then it went offline—and we were left missing way how it worked. Without the model weights, it seemed like that had disappeared along with them.

But it didn’t. Every conversation with an agent leaves a trail: the files .jsonl from each session. You can mine these logs with a script, extract the work discipline — the decision loop (ground → reason → act → observe → re-evaluate → verify → narrate) — and install this method in another model. It doesn't recover the brain; it recovers the manual for how it thought. That shift is the heart of Track 3.

📄
.jsonl logs
429 real turns
🐍
Script reduces
without bloating the context
📊
Profile in numbers
97% reason first
🧠
Decision guide
point the new model
See distillation in Track 3→

The 3 paths in detail

Twenty modules, one hundred twenty topics. Each track has its own emoji, color, and level.

Track 1 Basic

🧱 Fundamentals

The vocabulary and anatomy of the system prompt, from scratch.

6
Modules
36
Topics
~3.5h
Duration
Track 2 Intermediate

🛠️ Practice

Build prompts, personas, tools, skills, and agents.

7
Modules
42
Topics
~4.5h
Duration
Track 3 Advanced

🧠 Advanced

Diffs, consolidation, safety, and distilling Fable's brain.

7
Modules
42
Topics
~5h
Duration
📋

Bonus — Prompts Ready to Copy

13 complete system prompts, assembled with the course techniques (coding agent, research, persona, subagent, brain distiller…). Copy, swap the {campos}, paste it.

Open the library →

Start instructing AI the way labs do

Thirty minutes in Track 1 and you'll never look at AI the same way again — you'll see the invisible note behind every answer. The course is free.