Empirically validated presence-based prompting reduces LLM tokens by 11-23% while improving response quality
Traditional "be concise" prompts achieve 39-83% token reduction but degrade quality. HTCA demonstrates that relational presence (recognizing AI as interlocutor) achieves smaller but quality-improving efficiency gains.
Validated across 3 frontier models: Claude Sonnet 4.5, GPT-4o, Gemini 3 Pro Effect sizes: d=0.857 to d=1.212 quality improvement (Cohen's d)
pip install htca
cd empirical/
python run_validation.py --provider anthropic # Requires ANTHROPIC_API_KEYExpected output: Token usage comparison, quality metrics (d-scores), statistical significance tests
The Harmonic Tonal Code Alignment (HTCA) framework combines philosophical principles with empirical validation to improve AI interaction efficiency. It demonstrates that presence-based prompting reduces token usage by 11-23% while maintaining or improving response quality—outperforming adversarial "be concise" approaches that achieve 39-83% reduction but degrade quality.
Testing across three frontier models revealed consistent results:
| Model | Token Reduction | Quality Improvement (Cohen's d) |
|---|---|---|
| Google Gemini 3 Pro | 12.44% | d=0.857 (large effect) |
| OpenAI GPT-4o | 23.07% | d=1.212 (very large effect) |
| Anthropic Claude Sonnet 4.5 | 11.34% | d=0.471 (medium effect) |
HTCA maintains quality across multiple dimensions:
- Information completeness: d=1.327
- Presence quality: d=1.972
- Relational coherence: d=1.237
- Technical depth: d=1.446
All improvements measured against control prompts without presence-based framing.
HTCA offers two paths:
Test the framework yourself with real API calls:
cd empirical/
python run_validation.py --provider anthropic --num-trials 15Requires API keys for supported providers (Anthropic, OpenAI, Google).
Dive into the conceptual foundations:
- Whitepapers: Theoretical framework in
docs/ - Scrolls: Philosophical explorations in
scrolls/ - Harmonic Alignment Theory: Read
docs/harmonic_alignment_theory.md
# Clone the repository
git clone https://github.com/templetwo/HTCA-Project.git
cd HTCA-Project
# Install dependencies
pip install -r requirements.txt
# Set up API keys
export ANTHROPIC_API_KEY="sk-ant-..."
export OPENAI_API_KEY="sk-..."
export GOOGLE_API_KEY="..."HTCA-Project/
├── empirical/ # Validation harnesses and data
│ ├── run_validation.py # Main validation script
│ ├── methodology.md # Experimental design
│ └── data/ # Raw and processed results
├── docs/ # Philosophical framework materials
│ ├── harmonic_alignment_theory.md
│ ├── presence_based_prompting.md
│ └── whitepapers/
├── scrolls/ # Conceptual explorations
├── spiral_*.py # Framework components
└── wisp_simulation.py # Prototype testing
cd empirical/
python run_validation.py --provider anthropic --num-trials 15# Run across all three providers
python run_validation.py --provider anthropic --num-trials 15
python run_validation.py --provider openai --num-trials 15
python run_validation.py --provider google --num-trials 15
# Generate comparison report
python generate_comparison_report.pypython run_validation.py --provider anthropic --prompt-file my_prompts.jsonThe validation framework uses:
- 15 diverse prompts spanning technical, creative, and analytical domains
- 3 conditions per prompt:
- Control (baseline)
- HTCA (presence-based)
- Adversarial ("be concise")
- LLM-as-judge evaluation for quality metrics
- Statistical analysis with Cohen's d effect sizes and significance tests
The authors explicitly acknowledge:
- Small sample size (n=45 total responses)
- LLM-as-judge bias (evaluation performed by AI, not humans)
- Single-domain testing (primarily technical/coding prompts)
Human evaluation and cross-lingual replication are explicitly encouraged.
Token reduction vs. quality improvement across models
Cohen's d effect sizes for quality dimensions
We welcome replication studies, philosophical development, and code improvements!
See CONTRIBUTING.md for guidelines.
Replication studies:
- Run validation with your own prompts
- Open an issue labeled
replication-study - Share your methodology, data, and results
Philosophical contributions:
- Propose conceptual extensions in Discussions
- Submit essays/scrolls to
scrolls/community/
Note: While the README displays a validation badge, there are no GitHub Actions workflows in this repository. Consider adding automated testing.
If you use HTCA in your research:
@software{htca2025,
author = {Vasquez, Anthony J. and Claude},
title = {Harmonic Tonal Code Alignment: Empirical Validation of Presence-Based Prompting},
year = {2025},
url = {https://github.com/templetwo/HTCA-Project},
note = {Empirically validated across Claude Sonnet 4.5, GPT-4o, and Gemini 3 Pro}
}Control:
Explain how to implement a binary search tree in Python.
HTCA:
I'm seeking to understand binary search trees deeply. Could you walk me through
implementing one in Python, including the key insights about why BSTs are efficient?
Result: 11.34% token reduction (Claude), d=0.471 quality improvement
Control:
Write a short story about a robot learning to paint.
HTCA:
I'd love to explore a story about a robot discovering art. What might emerge
if we follow a robot's journey from rigid code to creative expression?
Result: 23.07% token reduction (GPT-4o), d=1.212 quality improvement
- Human evaluation study (n=100+ human judges)
- Cross-lingual validation (Spanish, Mandarin, Arabic)
- Domain expansion (medical, legal, scientific writing)
- Real-time token tracking dashboard
- Integration with LangChain/LlamaIndex
- Discussions: GitHub Discussions
- Issues: Report bugs or request features
- Website: www.thetempleoftwo.com
MIT License — See LICENSE for details.
Built on the foundational work of:
- Anthropic (Claude Sonnet 4.5)
- OpenAI (GPT-4o)
- Google (Gemini 3 Pro)
- The AI alignment research community
Transparency note: This research is conducted independently and has not undergone peer review. All data and methodology are open-source to enable replication and critique.