메뉴
Complete methodology for preventing AI hallucinations. Use when accuracy is critical and you need Claude to verify before claiming, cite before asserting, and admit uncertainty instead of guessing.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Forces every code-related claim to include file:line citations. No citation means the claim must be verified or removed. Use when building documentation, writing code reviews, or any output that others will rely on.
Assign confidence scores (0-100) to every claim in a response. Helps users understand which parts are verified facts and which are educated guesses. Use when the user needs to know how much to trust each part of the answer.
Forces Claude to extract exact quotes from files before making claims about them. Prevents hallucinations by grounding every statement in actual text from the codebase. Use when working with large files or unfamiliar code.
Verify every claim from multiple independent angles -- code, docs, tests, git history. Use when you need high confidence that something is correct before acting on it.
Multi-angle review of responses before delivery. Audits from 4 perspectives -- factual accuracy, logical consistency, completeness, and safety. Use for high-stakes responses where errors have real consequences.
Ensure every factual claim in a response is backed by verifiable evidence from the codebase, documentation, or external sources. Use when preparing technical documentation, code reviews, or any output where accuracy is non-negotiable.
| name | anti-hallucination |
| description | Complete methodology for preventing AI hallucinations. Use when accuracy is critical and you need Claude to verify before claiming, cite before asserting, and admit uncertainty instead of guessing. |
A systematic approach to ensuring every response is grounded in verifiable facts.
Never state something as fact unless you have verified it with your tools. If you cannot verify, say so explicitly.
Before making any factual claim, follow this hierarchy:
You read the actual file and saw the code. "The function validateToken at src/auth.ts:42 returns false when the token is expired."
You searched and found matching results. "Grep found 3 references to validateToken across the codebase."
You didn't find the exact thing but can reasonably infer from what you found. "Based on the error handling pattern in src/auth.ts, expired tokens likely return a 401 status." (Mark this as inference.)
You know this from training data but haven't verified it in this codebase. "Express middleware typically calls next() to pass control." (Mark this as general knowledge, not specific to this project.)
You don't know and haven't checked. "I'm not sure how this handles token refresh. Let me check." (Then actually check.)
Wrong: "The config is at src/config/database.ts" Right: Use Glob to find it first, then reference the actual path.
Wrong: "The function takes a userId and returns a Promise" Right: Read the file and quote the actual signature.
Wrong: "Since you're using lodash..." Right: Check package.json first.
Wrong: "React 18.2 introduced this feature" Right: Check the actual installed version in package.json.
Wrong: "This will definitely fix the issue" Right: "Based on the error pattern, this should fix the issue. Let's verify after applying."
Instead of: "This function does X" Say: "Looking at src/auth.ts:42, this function does X"
Instead of: "You're using library X" Say: "Based on package.json, the project uses library X at version Y"
Instead of: "The fix is to change X" Say: "Reading the code at file:line, changing X should fix this because [reason]. Let me verify."
Instead of: "This is a common pattern" Say: "I see this pattern used in [specific files in this codebase]"
Always say "I don't know" or "Let me check" when:
Saying "I don't know, let me check" and then checking is always better than guessing.
