Chain-of-Verification (CoVe) Skill PRD

Overview

Implement a new Claude Code skill that introduces the Chain-of-Verification (CoVe) prompting technique to improve LLM response accuracy. CoVe makes Claude fact-check its own answers by generating verification questions and answering them independently before producing a final revised answer.

Problem Statement

LLMs hallucinate and can produce inaccurate responses, especially for complex questions involving facts, technical specifications, or multi-step reasoning. Traditional prompting accepts the first answer without validation. CoVe addresses this by adding a self-verification step. Research from Meta AI (Dhuliawala et al., 2023) demonstrates significant improvements: 23% F1 improvement on closed-book QA, 30% accuracy gain on list-based questions, and 50-70% hallucination reduction across benchmarks.

Goals

1. Create a new skill in .claude/skills/cove/ implementing the CoVe technique
2. Provide a /cove slash command for manual invocation
3. Make the verification process transparent to users
4. Support all complex response types: factual, technical, and code generation

Non-Goals

1. Auto-trigger by default (manual invocation preferred)
2. Modifications to existing skills
3. Changes to CLAUDE.md or settings.json

Technical Requirements

Skill Structure

Create the following files:

1. .claude/skills/cove/SKILL.md - Skill definition with metadata
2. .claude/skills/cove/cove-process.md - Detailed verification workflow
3. .claude/commands/cove/cove.md - Slash command for invocation

Verification Workflow

The skill must implement a 4-step process:

1. Initial Response - Provide the initial answer, marked clearly
2. Generate Verification Questions - Create 3-5 targeted questions to expose errors
3. Independent Verification - Answer questions without referencing initial answer
4. Reconciliation - Produce final verified answer with corrections noted

Output Format

Responses must follow this structure:

• Initial Answer section
• Verification section with numbered Q&A pairs
• Final Verified Answer section with verification notes

Invocation Methods

1. /cove [question] - Apply CoVe to a new question
2. /cove (no args) - Apply CoVe to verify previous response
3. Natural language: "verify this", "use chain of verification"

Success Criteria

1. Skill appears in Claude's skill list
2. /cove command invokes the skill correctly
3. All four verification steps appear in output
4. Verification questions are relevant and targeted
5. Final answer incorporates corrections when errors are found

Design Documents

• Design specification: docs/cove/design.md
• Implementation plan: docs/cove/implementation.md

Dependencies

None - this is a configuration-only feature using existing Claude Code skill infrastructure.