# AoT-Skills
 
Atom of Thoughts reasoning framework as Claude skills. Based on ["Atom of Thoughts for Markov LLM Test-Time Scaling"](https://arxiv.org/abs/2502.12018) (Teng et al., NeurIPS 2025).
 
No MCP server needed — Claude internalizes the decomposition-contraction methodology and applies it natively.
 
## Skills
 
| Skill | File | Description |
|-------|------|-------------|
| `atom-of-thoughts` | `SKILL.yaml` | Full AoT: DAG decomposition, contraction, confidence scoring, verification, termination (depth 5) |
| `atom-of-thoughts-light` | `SKILL.yaml` | Fast variant: single cycle, depth 3, simplified confidence |
 
Prompt templates for decompose/contract/verify phases in `atom-of-thoughts/references/prompts.yaml`.
 
## Structure
 
```
skills/
├── atom-of-thoughts/
│   ├── aot.yaml
│   └── references/prompts.yaml
└── atom-of-thoughts-light/
    └── aot-light.yaml
```
 
## Why YAML
 
Token-efficient and LLM-friendly. Structured data parses cleaner than prose markdown for instruction-following.