An open API service indexing awesome lists of open source software.

https://github.com/snowmead/questionably-ultrathink

Questionably UltraThink your most complex tasks combining AoT (Atom of Thoughts) and CoV (Chain of Verification)
https://github.com/snowmead/questionably-ultrathink

agent-skills agents claude-code claude-code-plugin claude-code-skill

Last synced: 6 months ago
JSON representation

Questionably UltraThink your most complex tasks combining AoT (Atom of Thoughts) and CoV (Chain of Verification)

Awesome Lists containing this project

README

          

# Questionably UltraThink

Advanced reasoning plugin integrating Atom of Thoughts (AoT) and Chain of Verification (CoVe) frameworks.

## Installation

### Claude Code

```bash
# Add the marketplace
/plugin marketplace add snowmead/questionably-ultrathink

# Install the plugin
/plugin install questionably-ultrathink@snowmead-marketplace
```

### Other AI Agents (OpenCode, Cursor, Codex, etc.)

```bash
bunx skills add snowmead/questionably-ultrathink
```

This installs the skill to your agent's skill directory automatically.

## Usage

| Command | Description |
|---------|-------------|
| `/questionably-ultrathink` | Full reasoning pipeline (AoT + CoVe) |
| `/decompose` | Break down complex problems into atomic questions |
| `/verify` | Verify factual claims with independent questioning |

### Automatic Activation

The skill automatically activates when you use trigger phrases:

* "be thorough", "analyze carefully", "make sure this is right"
* "verify", "double-check", "are you sure"
* Complex multi-part questions
* Architecture or security decisions

## How It Works

1. **Atom of Thoughts (AoT)** - Decomposes complex problems into atomic sub-questions organized as a DAG
2. **Chain of Verification (CoVe)** - Verifies each answer through independent questioning with factored execution
3. **Per-Atom Folders** - Each atom has its own folder with `question.md`, `claims.md`, `answer.md`
4. **File-Existence State** - Pipeline state determined by which files exist (no mutable status fields)
5. **Factored Verification** - Verification questions answered by isolated verifiers with zero context about claims

### Architecture

```
┌─────────────────────────────────────────────────────────────────┐
│ User Query │
└─────────────────────────────────────────────────────────────────┘


┌─────────────────────────────────────────────────────────────────┐
│ Graph Generator │
│ Builds question DAG with dependencies │
│ Creates: atoms/{id}/question.md for each atom │
└─────────────────────────────────────────────────────────────────┘

┌──────────────────┼──────────────────┐
▼ ▼ ▼
┌─────────────────┐ ┌─────────────────┐ ┌─────────────────┐
│ Claim Generator │ │ Claim Generator │ │ Claim Generator │
│ (A1) │ │ (A2) │ │ (A3) │
│ [ISOLATED] │ │ [ISOLATED] │ │ [ISOLATED] │
│ Reads: │ │ Reads: │ │ Reads: │
│ question.md │ │ question.md │ │ question.md │
│ Writes: │ │ Writes: │ │ Writes: │
│ claims.md │ │ claims.md │ │ claims.md │
│ verifiers/*.md │ │ verifiers/*.md │ │ verifiers/*.md │
│ (question only)│ │ (question only)│ │ (question only)│
└─────────────────┘ └─────────────────┘ └─────────────────┘
│ │ │
▼ ▼ ▼
┌─────────────────────────────────────────────────────────────────┐
│ Factored Verification │
│ Verifier reads pre-created file containing ONLY the question │
│ Verifier has ZERO context about the original claim │
│ Overwrites: atoms/{id}/verifiers/{N}.md with answer │
└─────────────────────────────────────────────────────────────────┘
│ │ │
▼ ▼ ▼
┌─────────────────┐ ┌─────────────────┐ ┌─────────────────┐
│ Verifier │ │ Verifier │ │ Verifier │
│ "What is X?" │ │ "When was Y?" │ │ "How does Z?" │
│ [NO CONTEXT] │ │ [NO CONTEXT] │ │ [NO CONTEXT] │
└─────────────────┘ └─────────────────┘ └─────────────────┘
│ │ │
└──────────────────┼──────────────────┘

┌─────────────────────────────────────────────────────────────────┐
│ Verification Maintainer │
│ Reads: claims.md + verifiers/*.md │
│ Cross-checks claims vs independent verifier answers │
│ Marks: VERIFIED | REVISED | REFUTED | UNCERTAIN │
│ Creates: answer.md with verification trace │
└─────────────────────────────────────────────────────────────────┘


┌─────────────────────────────────────────────────────────────────┐
│ Graph Maintainer │
│ Contracts solved answers INTO dependent questions │
│ Updates question.md: "Given A1=X, A2=Y, what is...?" │
└─────────────────────────────────────────────────────────────────┘


[Repeat for each level]


┌─────────────────────────────────────────────────────────────────┐
│ FINAL Answer │
│ Read from atoms/FINAL/answer.md │
└─────────────────────────────────────────────────────────────────┘
```

### File Structure

```
.questionably-ultrathink/{session-id}/
├── metadata.md # DAG structure (immutable)
└── atoms/
├── A1/
│ ├── question.md # Created by graph-generator
│ ├── claims.md # Created by claim-generator
│ ├── answer.md # Created by verification-maintainer
│ └── verifiers/
│ ├── 1.md # Pre-created by claim-generator (question only)
│ │ # Completed by verifier (adds answer)
│ └── 2.md
└── FINAL/
└── ...
```

**State = File Existence + Content:**

* `question.md` exists → atom created
* `claims.md` exists → claims generated
* `verifiers/{N}.md` exists with question only → pre-created, ready for verifier
* `verifiers/{N}.md` has "# Answer" section → verification question answered
* `answer.md` exists → verification complete

**Key insight:** Factored verification from the CoVe paper. Each verifier reads a pre-created file containing ONLY the verification question - no claim text. The verifier has ZERO knowledge of the original claim. This prevents bias - when verifiers don't know the "expected" answer, they can't be biased toward confirming it.

### Rigor Levels

| Level | Re-solve Trigger | Use Case |
|-------|------------------|----------|
| **Standard** | Never (single pass) | Most questions |
| **Thorough** | LOW confidence atoms | Important decisions |
| **High-Stakes** | Below HIGH confidence | Security, architecture, production |

## Optional: Parallel.ai MCP Integration

The plugin includes optional MCP servers for enhanced web search during verification:

* `parallel-search` - Optimized fact-checking searches
* `parallel-task` - Deep research capabilities

**Setup:** Run `/mcp` in Claude Code and authenticate with Parallel.ai.

**Fallback:** Works fully without MCP, using native `WebSearch` and `WebFetch` tools.

## References

* [Atom of Thoughts Paper](https://arxiv.org/abs/2502.12018) - HKUST, 2025
* [Chain-of-Verification Paper](https://arxiv.org/abs/2309.11495) - Meta AI, 2023