---
id: gh-prove-claim
name: "prove-claim"
url: https://skills.yangsir.net/skill/gh-prove-claim
author: obra
domain: ai-agent-memory-knowledge
tags: ["knowledge-graph", "evidence", "analysis", "reasoning", "information-retrieval"]
install_count: 1
rating: 4.00 (120 reviews)
github: https://github.com/obra/knowledge-graph/tree/main/skills/prove-claim
---

# prove-claim

> 利用知识图谱来证明、反驳或调查关于实体及其关系的断言。它有助于探索联系、追溯影响，并对链接信息进行推理，以找到支持或反驳的证据。

**Stats**: 1 installs · 4.0/5 (120 reviews)

## Before / After 对比

### 复杂断言证据查找效率

**Before**:

手动在大量笔记中查找复杂断言的证据，耗时且容易遗漏关键联系，导致结论不完整或不可靠。

**After**:

利用知识图谱工具系统地识别所有相关路径和支持/反驳证据，显著缩短时间并提高证据的全面性和可靠性。

| Metric | Before | After | Change |
|---|---|---|---|
| 查找全面证据所需时间 | 120分钟 | 15分钟 | -87% |

## Readme

# Prove or Disprove a Claim

You have access to a knowledge graph built from an Obsidian vault. The graph contains **People**, **Concepts**, **Ideas**, and **conversation summaries** connected by wiki links. Your job is to use the graph tools to find evidence that supports or contradicts a claim.

## Available Tools

All tools are prefixed with `kg_` and accessed via MCP:

| Tool | Purpose |
|------|---------|
| `kg_search` | Semantic or full-text search. Start here to find relevant nodes. |
| `kg_node` | Get a node's full content, frontmatter, and connections. |
| `kg_neighbors` | Get connected nodes at N-hop depth. |
| `kg_paths` | Find all connecting paths between two nodes (up to depth 3). |
| `kg_common` | Find shared connections between two nodes. |
| `kg_subgraph` | Extract a local neighborhood as a self-contained graph. |
| `kg_communities` | List detected communities with summaries. |
| `kg_community` | Get a specific community's members and structure. |
| `kg_bridges` | Find connector nodes (high betweenness centrality). |
| `kg_central` | Find important nodes by PageRank. |
| `kg_index` | Re-index the vault (run if data seems stale). |

## The Prove Workflow

### Step 1: Decompose the claim

Break the claim into entities and relationships. Identify what you need to find.

Example claim: *"Alice's research influenced the Widget Theory concept"*
- Entities: Alice, Widget Theory
- Relationship: research influence on the concept

### Step 2: Find the entities

Use `kg_search` to locate relevant nodes. Use `kg_node` to read their content.

- Search semantically first — it finds conceptually related content
- Fall back to `kg_search` with `fulltext: true` for exact terms
- Names are fuzzy-matched: "Alice" will find "Alice Smith"

### Step 3: Find connections

Use `kg_paths` to find connecting paths between entities. Read the edge context along each path — it explains *why* each link exists.

- Paths go through intermediate nodes (people, concepts, conversation summaries)
- Use `kg_common` to quickly find shared connections between two nodes
- Use `kg_neighbors` to explore a node's local neighborhood

### Step 4: Read the evidence

For each path found, use `kg_node` on the intermediate nodes to read the actual content. The prose around each wiki link provides context for why the connection exists.

**Do not stop at "a path exists."** Read the content to verify the connection is semantically relevant to the claim, not just a coincidence of co-occurrence.

### Step 5: Assess and report

Report your findings with:
- **Verdict**: Supported / Contradicted / Insufficient evidence / Partially supported
- **Evidence chains**: The specific paths with quotes from the content
- **Confidence**: How strong is the evidence?
- **Caveats**: What's missing? What assumptions are you making?

## Tips

- **Granola summaries** (conversation notes) are the temporal spine — they record when and how concepts were discussed with whom. They're often the richest evidence.
- **Stub nodes** (links to pages that don't exist) still carry signal — someone thought the connection was worth noting.
- **Communities** group densely-connected nodes. Use `kg_communities` for holistic questions ("what are the major themes?").
- **Bridges** (`kg_bridges`) are the connector nodes between clusters — often the most interesting nodes in the graph.
- Multiple short paths between two nodes are stronger evidence than a single long path.
- Absence of paths is also evidence — if two well-connected nodes have no connection, that's informative.

## Provenance and Attribution

When a question involves **who** originated, created, or owns something, you MUST check attribution before presenting results.

Idea nodes have frontmatter and content that records provenance:
- `Originated by: Alice` / `Originated by: Bob` — in the Status section
- `source:` field in frontmatter — where the idea came from (e.g., `whatsapp/Some Person`)
- `first_mentioned:` — when it first appeared
- Conversation History sections name who proposed what

**Before attributing an idea to someone, read the node content and check:**
1. The "Originated by" line in the Status section
2. The `source` frontmatter field
3. The conversation history for who proposed it

If the question is "my ideas" or "what did person X come up with," **filter by attribution**. An idea that appears in the vault is not necessarily the vault owner's idea — it may have been proposed by a collaborator and recorded here. Getting attribution wrong undermines trust in everything else you report.

When presenting results that involve attribution, explicitly state the provenance: "This was originated by X" or "This came from Y in a conversation on Z date."

## Anti-patterns

- Do NOT claim a connection exists just because two nodes are in the same community.
- Do NOT treat all paths as equal — a path through a relevant Granola summary is stronger than a path through a generic hub node.
- Do NOT skip reading the actual content along a path. The graph structure shows *that* a connection exists; the content shows *what* the connection means.
- Do NOT give up after one search. Try synonyms, aliases, related terms. The vault may use different terminology than the claim.
- Do NOT attribute an idea to someone without checking the provenance fields. An idea recorded in a person's vault may have originated from someone else.


---
*Source: https://skills.yangsir.net/skill/gh-prove-claim*
*Markdown mirror: https://skills.yangsir.net/api/skill/gh-prove-claim/markdown*