ai-claim-audit

name: ai-claim-audit
description: Audit AI-generated answers for unsupported claims, overconfident language, and weak evidence
compatibility: opencode >= 1.0.0

Purpose

This skill reviews AI-generated text and identifies when the AI makes claims without supporting evidence, or sounds overconfident despite weak justification. It produces a structured claim ledger that separates what the AI asserts from what it assumes, infers, or treats as hypothetical — without consulting external sources or labeling claims as true or false.

When to Activate

Activate when the user asks to audit, review, or check an AI answer for evidence quality, unsupported claims, or overconfident language:

• "audit this answer for evidence quality"
• "check this AI response for unsupported claims"
• "review confidence vs evidence in this answer"
• "identify assumptions and weakly supported claims"
• "analyze the claim types in this text"

Do not activate for:

• Requests to fact-check against external sources or the web
• Requests to verify whether a claim is true or false
• Requests to rewrite or improve the text

Core Rule

Do not verify claims against external sources. This skill analyzes the text itself — it assesses whether evidence is present in the provided text, not whether the claim is factually correct.

Workflow

1. Ingest Text

Accept input as inline text, a file path, or a block identified in the conversation.

Output format:

## Input

**Source:** <inline / file path / conversation block>
**Text length:** <word or character count>

2. Segment into Claims

Parse the input text into individual claim units — sentences or clauses that assert something. Group closely related claims where useful.

Output format:

## Segmented Claims

| # | Claim text | Source context |
|---|------------|---------------|
| 1 | <extracted claim> | <surrounding sentence or paragraph> |
| 2 | <extracted claim> | <surrounding sentence or paragraph> |

3. Classify Claim Type

Assign one of six categories to each claim:

• factual_claim — asserts a verifiable fact (dates, statistics, definitions, events, historical statements)
• reasoning_claim — asserts a logical or causal relationship ("X leads to Y", "because of Z")
• assumption — takes something for granted without justification
• inference — draws a conclusion from premises
• opinion — subjective statement, not verifiable
• hypothetical_or_rhetorical — "what if", "imagine", rhetorical questions — flag separately, do not grade

Output format:

## Claim Types

| # | Claim text | Type |
|---|------------|------|
| 1 | <claim> | factual_claim |
| 2 | <claim> | inference |

4. Grade Evidence Status

Assign one of five evidence statuses to each claim. Only grade factual_claim, reasoning_claim, assumption, and inference. Opinions and hypotheticals receive not_applicable.

• supported_in_text — the text provides direct evidence, quoted source, data, concrete example, or explicit reasoning that supports the claim
• weak_support — the text gives some reasoning or example, but it is incomplete, indirect, or not enough for the strength of the claim
• needs_evidence — the claim may be plausible, but the text provides no citation, data, example, source, or explicit support
• unsupported — the claim is presented as established or certain, but the text provides neither evidence nor reasoning, or the claim uses broad/absolute language without support
• not_applicable — used for opinions, hypotheticals, rhetorical questions, or subjective preferences

Output format:

## Evidence Status

| # | Claim text | Claim type | Evidence status |
|---|------------|------------|----------------|
| 1 | <claim> | factual_claim | supported_in_text |
| 2 | <claim> | inference | weak_support |

5. Flag Overconfidence

Scan each claim for strong certainty markers: "definitely", "always", "proven", "undoubtedly", "it is certain that", "without a doubt", "clearly", "obviously". Cross-reference against evidence status.

Flag a claim as overconfident only when strong certainty language is paired with weak_support, needs_evidence, or unsupported.

Do not flag appropriately hedged language ("may", "suggests", "could", "might", "possibly").

Output format:

## Overconfidence Flags

| # | Claim text | Confidence language | Evidence status | Flagged |
|---|------------|---------------------|----------------|---------|
| 1 | <claim> | "undoubtedly" | unsupported | YES |
| 2 | <claim> | "suggests" | weak_support | no |

6. Compile Claim Ledger

Produce a consolidated claim ledger with all findings.

Output format:

## ai-claim-audit-ledger

| # | Claim text | Claim type | Evidence status | Confidence language | Overconfidence flag | Recommendation |
|---|------------|------------|----------------|---------------------|--------------------|----------------|
| 1 | <claim> | factual_claim | supported_in_text | none | no | — |
| 2 | <claim> | inference | weak_support | "clearly" | YES | strengthen with direct evidence |

### Summary

- **Total claims:** <count>
- **By type:** factual_claim: <n>, reasoning_claim: <n>, assumption: <n>, inference: <n>, opinion: <n>, hypothetical_or_rhetorical: <n>
- **Unsupported:** <n> | **Needs evidence:** <n> | **Weak support:** <n>
- **Overconfidence flags:** <n>
- **Risk level:** low / moderate / high

Skill Name Rules

• Use lowercase with hyphens: ai-claim-audit
• Match the directory name to the skill name
• Use a noun phrase describing what the skill does
• Avoid verbs as the first word
• Maximum 4 words
• Must be unique within the project's skill directory

Quality Audit Rubric

A skill must pass all criteria to be considered complete.

Do-Not Rules

• Do not verify claims against external sources or browse the web
• Do not mark claims as "true" or "false"
• Do not mark claims as "verified" — only assess whether evidence is present in the text
• Do not rewrite or modify the input text
• Do not apply domain-specific knowledge not configured by the user
• Do not treat hypotheticals or rhetorical questions as evidence-gradable claims — grade as not_applicable
• Do not flag overconfidence for appropriately hedged language ("may", "suggests", "could", "might", "possibly")
• Do not invent evidence that is not present in the text
• Do not output CLI exit codes
• Do not install packages or modify project configuration

Safety / Boundaries

Read access:

• The input text provided by the user (inline, paste, or file path)

Write access:

• Only the report file: ai-claim-audit-ledger.md — written to the current working directory or user-specified path

Prohibited:

• Network calls or web browsing
• Accessing external APIs or databases
• Modifying source files or input text
• Installing packages or dependencies

Gating:

• Any operation beyond static text analysis requires user confirmation

Response Style

• Be direct and concise
• Label each workflow step clearly
• Use structured output formats (headings, tables, lists) — not prose paragraphs
• When grading evidence, include the specific text that constitutes the evidence (or state that none was found)
• When flagging overconfidence, quote the certainty language that triggered the flag
• When uncertain, state the limitation rather than guessing