cove-prompt-design
Skill67 starsChanged 55 days agoMITView on GitHub
Explains Chain of Verification (CoVe) prompt design — a 4-step pattern separating generation from independent factual verification. Use when designing prompts that require factual accuracy, reducing hallucinations, checking technical standards or APIs, or producing step-by-step procedures where subtle errors are costly.
In this repository41 other files it gives its agents.
- AGENTS.mdAGENTS.md
- rules/AGENTS.mdAGENTS.md
- .claude/CLAUDE.mdCLAUDE.md
- research/CLAUDE.mdCLAUDE.md
- .cursor/rules/backlog-before-work.mdcCursor rule
- agent-browserSkill
- beadsSkill
- cccSkill
- copy-editingSkill
- daily-releasesSkill
- graph-engineeringSkill
- rebaseSkill
- receiving-pr-reviewsSkill
- topic-specialistSkill
- workshop-question-framingSkill
- agent-browserSkill
- boilSkill
- commit-stagedSkill
- complete-milestoneSkill
- create-merge-request-changelogSkill
- create-milestoneSkill
- design-anti-patternsSkill
- evaluate-sdlc-layersSkill
- example-argument-substitutionSkill
- external-pattern-integratorSkill
- fact-checkSkill
- ghSkill
- git-history-reconSkill
- group-items-to-milestoneSkill
- knowledge-explorerSkill
- linear-walkthroughSkill
- modern-gitSkill
- prepare-walkthrough-presentationSkill
- readme-badgerSkill
- refresh-researchSkill
- research-curatorSkill
- session-historianSkill
- seven-prompt-content-engineSkill
- skill-research-processSkill
- start-milestoneSkill
- universe-of-thoughtsSkill
The file
We split this skill into the rules it gives an agent
5 rules. About 0 are common, 5 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈5
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 33
Use CoVe when any of the following are true:
- Distinctive: line 49
Do not use CoVe for:
- Distinctive: line 165
If you want concise results, explicitly constrain output:
- Distinctive: line 182
Use CoVe to reduce hallucinations and increase reliability by:
- Distinctive: line 188
If the task can be wrong in subtle ways, CoVe is appropriate.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Chain of Verification (CoVe) for Prompt Design
This skill documents what CoVe is, when to use it, and how to apply it correctly in prompt design.
CoVe is a prompt design pattern that improves factual reliability by separating generation from verification. The model first produces an answer, then independently verifies that answer through structured checks, and finally produces a corrected or confirmed result.
This is not chain of thought exposure. The verification steps are explicit instructions, not hidden reasoning.
What CoVe Is
Chain of Verification (CoVe) is a structured prompting approach with three phases:
- Initial answer generation
- Independent verification questions
- Final validated answer
The key idea is that the model should not assume its first answer is correct. It must actively test it.
When to Use CoVe
Use CoVe when any of the following are true:
- The task requires factual accuracy (dates, specs, standards, APIs).
- Hallucinations would be costly or misleading.
- The answer depends on multiple independent facts.
- The model may rely on weak priors or pattern completion.
- You want the model to challenge its own output.
Concrete examples:
- Explaining technical standards or protocols.
- Summarizing regulations or compliance requirements.
- Producing step by step procedures with real world consequences.
- Comparing versions, limits, or constraints.
- Answering questions that look simple but are easy to get subtly wrong.
Do not use CoVe for:
- Creative writing.
- Brainstorming.
- Open ended ideation.
- Pure opinion or preference questions.
Core CoVe Structure
A minimal CoVe prompt has this structure:
Step 1: Produce an initial answer.
Step 2: Generate verification questions that would test the answer.
Step 3: Answer each verification question independently.
Step 4: Revise or confirm the original answer based on verification.The verification questions must be phrased so they can falsify the answer.
Verification Question Design Rules
Good verification questions:
- Are independent of the original wording.
- Target factual claims, not style.
- Can be answered without referring to the initial answer.
- Are specific and falsifiable.
Bad verification questions:
- "Is the answer correct?"
- "Does this look right?"
- "Explain why this is good."
Good examples:
- "What is the maximum allowed value according to the spec?"
- "Does the standard permit this behavior?"
- "Is this API behavior documented or inferred?"
Example: Without CoVe
Explain how HTTP caching works.Risk:
- Model may hallucinate headers, defaults, or behaviors.
Example: With CoVe
Task: Explain how HTTP caching works.
Step 1: Provide an initial explanation.
Step 2: List verification questions that check factual correctness
(e.g. cache control headers, default behaviors, validation rules).
Step 3: Answer each verification question independently using known standards.
Step 4: Produce a final corrected explanation.CoVe vs Chain of Thought
Important distinction:
- Chain of Thought exposes reasoning.
- CoVe enforces verification.
You are not asking the model to show hidden reasoning. You are asking it to run checks.
Verification outputs may be shown or summarized, depending on your needs.
Failure Modes to Watch For
Common mistakes when using CoVe:
- Verification questions that simply restate the answer.
- Letting the model reuse the same phrasing across steps.
- Skipping revision even when verification reveals uncertainty.
- Using CoVe when accuracy is not required (adds unnecessary verbosity).
Practical Prompt Template
You will complete this task using Chain of Verification.
1. Produce an initial answer.
2. Generate 3 to 6 verification questions that could falsify the answer.
3. Answer each verification question independently.
4. If any answer reveals an error or uncertainty, revise the original answer.
5. Output only the final revised answer, followed by a short confidence note.Output Control
If you want concise results, explicitly constrain output:
- Hide intermediate verification steps.
- Summarize verification findings.
- Emit only the final answer unless issues were found.
Example:
Perform verification internally.
Only show the final answer and list any corrected assumptions.Summary
Use CoVe to reduce hallucinations and increase reliability by:
- Separating generation from verification.
- Forcing independent factual checks.
- Revising answers based on evidence, not confidence.
If the task can be wrong in subtle ways, CoVe is appropriate.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

