 TIC — Output |
 Non-TIC — Output |
Both images generated with Z-Image-Base 8bit — same prompt idea, different reasoning depth.
TIC (Taken In Consideration) is a cascade-thinking methodology implemented as a skill for AI agents. It gives LLMs an additional layer of systematic reasoning focused on exhaustive consideration of all possible cases, consequences, and edge scenarios before any execution begins.
The core problem it addresses: LLMs lack innate spatial and physical reasoning. They default to happy-path thinking and frequently miss edge cases, spatial contradictions, physical consequences, and cascading effects of decisions. TIC forces the model to decompose a task atomically, cascade every element through all possible scenarios, cross-check intersections, and stress-test against reality — before producing any output.
This repository contains initial proof-of-concept tests comparing outputs with and without TIC across three domains:
| Domain | Files | What to compare |
|---|---|---|
| Image Generation | ProofOfConcept/ImageGeneration/ |
Prompts + generated images — spatial coherence, perspective correctness, physical plausibility |
| Creative Writing | ProofOfConcept/Writing/ |
Two short stories from the same prompt — narrative consistency, cause-and-effect logic |
| Coding / System Design | ProofOfConcept/Coding/ |
Architecture designs for a restaurant reservation system — edge case coverage, concurrency handling |
⚠️ Note: All test prompts and responses are written in Italian. Use your AI agent of choice to translate and understand the results.
All tests were run under identical conditions to ensure fair comparison:
- Harness: pi — the same coding agent environment for both runs
- Model: Qwen 3.6 27B (fine-tuned variant from Hugging Face) — same model, same session config
- Variables: The only difference is whether or not the TIC skill was loaded and triggered
The number of tests is small and more are needed for statistical significance. However, even in these initial analyses:
- Spatial coherence in image generation prompts is significantly higher with TIC — the model correctly reasons about perspective, lighting direction, anatomical orientation, and environmental context before writing the prompt
- Writing outputs show stronger cause-and-effect chains and fewer narrative inconsistencies
- Coding designs cover substantially more edge cases (concurrency, failure modes, boundary conditions) without being explicitly asked
The difference is most visible in image generation, where TIC's reality-checking phase directly translates into prompts that produce spatially consistent results.
TIC is distributed as a skill. Load SKILL.md into your agent and trigger it with phrases like:
"usa TIC""think with TIC""considerazione TIC"
It will run through the 5-phase cascade (Decompose → Cascade → Cross-Check → Stress-Test → Identify Gaps) before producing any deliverable.
TBD