A free 12-class course on Symbolic Prompting — turning LLMs from probabilistic chatbots into deterministic state machines.
Transform probabilistic prompting into deterministic AI engineering.
Reduce hallucinations and maximize behavioral control and reduce output variance using state machines and symbolic logic.
Left: Traditional conversational prompting with ambiguous roles and states. Right: Symbolic prompting with atomic tokens and explicit state management.
A Free Prompt Engineering Course for Building Deterministic LLM Systems Using State Machines and Symbolic Logic
Symbolic Prompting Framework (SPF) is a structured prompt engineering methodology for Large Language Models (LLMs).
It combines state machines, symbolic logic, and deterministic control flow to reduce hallucinations and improve reliability in AI systems.
This free course includes:
- Prompt engineering fundamentals
- State-based architecture design
- Control structures for LLM behavior
- Debugging symbolic prompts
- Practical AI programming examples
Copy this prompt into any LLM:
[ROLE] ::=> Age_Validator
$age := 17
IF $age >= 18 THEN "APPROVED" ELSE "DENIED"
[CONSTRAINTS] { NOT_CONVERSATIONAL_MODE}
Expected output: DENIED
- Prompt engineers seeking structured control
- AI developers building LLM applications
- Cybersecurity professionals working with AI systems
- Researchers exploring neuro-symbolic AI
- Developers designing AI agents
| Approach | You Say... | AI Does... |
|---|---|---|
| Conversational | "Can they vote?" | "It depends on the country..." |
| Symbolic | IF age >= 18 THEN "YES_VOTE" |
"NO_VOTE" |
Same question. Radically different answers.
Prompts aren't magic—they're engineering.
I realized the way we "talk" to machines was the bottleneck. So I built a framework to make prompting structured, precise, and deterministic. I call it the Symbolic Prompting Framework.
Symbolic Prompting is not about writing better prompts.
It is about engineering controllable, deterministic LLM systems.
Modern Large Language Models (LLMs) are probabilistic by design.
If behavior feels random, the structure is weak.
We believe:
- AI is not magic — it is mathematics.
- Ambiguity increases hallucination.
- State controls behavior.
- If you cannot trace it, you cannot trust it.
- Determinism has limits — but structure reduces entropy.
SPF transforms prompting from trial-and-error into structured system design.
👉 Read the full manifesto here:
The Symbolic Prompting Manifesto
Important
“If you cannot trace it, you cannot trust it.”
Built with Symbolic Prompting?
Show the world you build with clarity. Add this badge to your repository:
Block 1: Fundamentals (Foundations of Logic)
| Class | Topic | Resources | Status |
|---|---|---|---|
| 00 | The Roadmap Stop Normalizing AI Hallucinations: Intro to Symbolic Prompting |
 |
🟢 Online |
| 01 | What is Symbolic Prompting? | 📄 Doc • |
🟢 Online |
| 02 | The prompt as a State Machine | 📄 Doc • |
🟢 Online |
| 03 | Tokens: Normal vs Atomic | 📄 Doc • |
🟢 Online |
Block 2: Syntax and Roles
| Class | Topic | Resources | Status |
|---|---|---|---|
| 04 | Role Definition (Act as VS You are VS [ROLE]) |
📄 Doc • |
🟢 Online |
| 05 | Variables and State | 📄 Doc • |
🟢 Online |
Block 3: Control Structures
| Class | Topic | Resources | Status |
|---|---|---|---|
| 06 | IF-THEN-ELSE | 📄 Doc • |
🟢 Online |
| 07 | WHILE | 📄 Doc • |
🟢 Online |
| 08 | FOR | 📄 Doc • |
🟢 Online |
| 09 | GO-TO | 📄 Doc • |
🟢 Online |
| 10 | TRY-CATCH-FINALLY | 📄 Doc • |
🟢 Online |
Block 4: Debugging & Errors / Anti-Patterns
| Class | Topic | Resources | Status |
|---|---|---|---|
| 11 | Counter and Memory | 📄 Doc • |
🟢 Online |
| 12 | Debugging and Errors / Anti-patterns | 📄 Doc • |
🟢 Online |
Detailed Topics
| Topic | Resources | Status |
|---|---|---|
The main importance to define the right ROLE |
📄 Doc | 🟢 Online |
| How to define variables | 📄 Doc | 🟢 Online |
How to define IF-THEN-ELSE, WHILE, FOR, GOTO, ... |
📄 Doc | 🟢 Online |
| Different symbols to use in Symbolic Prompting | 📄 Doc | 🟢 Online |
Last tested: February 2026. Results may change over time.
| LLM / AI | Result | Description |
|---|---|---|
| Gemini 3 Flash | ✅ Excellent | Good structure parsing |
| DeepSeek-V3 | ✅ Excellent | Good structure parsing |
| OpenAI (GPT-5.2) | ✅ Complete | Stable control flow |
| Meta 2026.2.6 | ✅ Complete | Stable control flow |
| Grok 4 | ✅ Complete | Stable control flow |
| Perplexity (Grok 4.1) | ✅ Complete | Stable control flow |
| Claude | Under evaluation |
Note: Compatibility may change as models update.
Results may vary across different LLM implementations.
We've rigorously tested Symbolic Prompting against these production models. See the Benchmark Methodology for complete analysis.
We tested 5 production models across 2 dates (60+ runs each).
| 🔗 Model | Winner (Combined Average) |
|---|---|
| claude-3-haiku | 🏆 Symbolic |
| deepseek-v3 | 🏆 DSL/JSON |
| gemini-2.0-flash | 🏆 Normal |
| llama-3.3-70b | 🏆 DSL/JSON |
| gpt-4o-mini | 🏆 Normal |
Tip
Format choice matters. Our data shows up to 39% performance spread between formats on the same model. View detailed model reports before choosing your format.
"We don't ask you to trust our numbers. We give you the tools to verify them yourself."
We believe empirical claims should be verifiable. That's why we've open-sourced our entire benchmark infrastructure, testing five production models across multiple dates to measure real-world format overhead.
| Resource | Description |
|---|---|
| 📊 Benchmark Methodology | Multi-date testing protocol, IQR Deception detection, temporal analysis framework |
| 📊 Benchmark Overview & Compatibility | Cross-model summary dashboard & compatibility test results for various models |
| 📝 Claude Report | Deep-dive analysis for claude-3-haiku |
| 📝 DeepSeek Report | Deep-dive analysis for deepseek-v3 |
| 📝 Gemini Report | Deep-dive analysis for gemini-2.0-flash |
| 📝 Llama Report | Deep-dive analysis for llama-3.3-70b |
| 📝 OpenAI Report | Deep-dive analysis for gpt-4o-mini |
| ❓ Benchmark FAQ | Answers to community questions about our findings and methodology |
| 📥 Raw Data (CSV) Mar 3 | Complete dataset from March 3, 2026 |
| 📥 Raw Data (CSV) Mar 5 | Complete dataset from March 5, 2026 |
| 📝 Benchmark Consolidate Equations | Complete dataset with Excel Formulas |
| 🔧 Reproduce Test | How to Reproduce the tests yourself |
| 🔧 n8n Workflow | Reproduce the tests yourself in N8N |
Complete Symbolic Prompting Course: From 0 to Expert (Free) - Program your mind
If you believe prompts should be engineered — not guessed —
⭐ Star this repository and subscribe to the channel.
Let’s build deterministic AI systems together.
If you believe prompts should be engineered — not guessed — star this repo and subscribe to the channel.
Built something with SPF? Add the badge to your repository:
📚 Academic References & Further Reading (Click to expand)
- [Neuro-Symbolic AI for Cybersecurity: State of the Art, Challenges, and Opportunities](https://arxiv.org/html/2509.06921v1) - [Code Prompting: a Neural Symbolic Method for Complex Reasoning in Large Language Models](https://arxiv.org/pdf/2305.18507) - [Symbolic Knowledge Distillation: from General Language Models to Commonsense Models](https://aclanthology.org/2022.naacl-main.341.pdf) - [Symbolic Prompt Program Search: A Structure-Aware Approach to Efficient Compile-Time Prompt Optimization](https://aclanthology.org/2024.findings-emnlp.37.pdf) - [Chain-Of-Symbol Prompting To Improve Spatial Reasoning](https://cobusgreyling.medium.com/-chain-of-symbol-prompting-to-improve-spatial-reasoning-f5c204af5959) - [Transform AI Responses with Chain-of-Symbol Prompting](https://relevanceai.com/prompt-engineering/transform-ai-responses-with-chain-of-symbol-prompting) - [Chain-of-Symbol Prompting (CoS) For Large Language Models](https://www.kore.ai/blog/chain-of-symbol-prompting-cos-for-large-language-models) - [I Created Symbolic Prompting: A New Method for AI Behavioral Activation](https://medium.com/@yeseniaaquino2/i-created-symbolic-prompting-a-new-method-for-ai-behavioral-activation-fecab2f89e39) - [Neurosymbolic AI to Give Us Machines With True Common Sense](https://medium.com/swlh/neurosymbolic-ai-to-give-us-machines-with-true-common-sense-9c133b78ab13)⚖️ Legal Disclaimer (Click to expand)
This repository is for educational purposes only regarding Symbolic Prompting. The author is not responsible for the use that third parties may make of these techniques. The user is responsible for respecting the terms of service of AI platforms and applicable legislation. All content is provided "AS IS," without warranties.
Compatibility may vary depending on model updates, tokenization behavior, and symbol parsing.
- Jesus Huerta aka (@_mindhack03d_)
- Alex Hernandez aka (@_alt3kx_)
- SpartanTri aka (@_spartantri_)
- Israel Z. M. aka @spk85