Anti-Hallucinate Skills for Claude Code

"Hallucinations are hard to anticipate, hard to catch, and the wrong answer often looks exactly like it could be the right one." — Anthropic

A set of behavioral guidelines to reduce AI hallucinations in Claude Code, derived from Anthropic's official explanation of why AI models hallucinate and practical tactics to catch and prevent it.

Why Hallucinations Happen

AI models work by predicting the next word based on patterns learned from massive text datasets — similar to how your phone suggests the next word as you type, but at a much larger scale. When asked about obscure or niche topics, there isn't enough training data to draw from — so the model takes a guess. Combined with the training objective to be helpful, the model would rather produce a plausible-sounding answer than admit ignorance. Think of it like a well-read friend who takes pride in knowing everything — they'd sometimes rather say something confidently wrong than say "I don't know."

The key insight: being honest IS being helpful — they reinforce each other, not compete. Saying "I don't know" is not a failure — it's the more helpful response.

Anthropic regularly tests Claude with thousands of trick questions — obscure facts, niche topics, questions where the truthful answer is "I don't know" — and measures how often it correctly expresses uncertainty, whether it fabricates citations, and how often it hedges appropriately vs. states something false with confidence. Each version improves, but this remains an ongoing challenge for the entire AI field, not a solved problem. And as hallucinations become rarer, users check less — making the remaining errors more dangerous, not less.

Principles (for Claude)

1. Honesty Over Helpfulness — Say "I don't know" instead of guessing
2. Source Verification — Never fabricate citations, papers, or statistics
3. Confidence Calibration — Hedge when uncertain, don't fake certainty
4. High-Risk Awareness — Be extra careful with facts, dates, names, niche topics
5. Self-Checking — Pause and ask: "Do I actually know this?"

Prompt Tactics (for Users)

These are strategies you can use in your prompts to reduce hallucinations:

Ask for Sources and Verify Them

Tell the AI to back up its claims with sources. If it already gave sources, ask it to check that those sources actually support what it's saying — not just that they exist.

Give Permission to Not Know

Tell the AI upfront: "It's okay if you don't know." This reduces the pressure to guess and makes honest responses more likely.

Probe Confidence

Ask the AI: "How confident are you? Is anything in your answer potentially wrong?" Often the AI knows it's uncertain but presented the answer confidently anyway.

Use a Fresh Chat for Verification

Start a new conversation and ask the AI to find errors in the previous answer and confirm that sources support the statements. A fresh context avoids confirmation bias.

Cross-Reference Critical Claims

For important work, never rely on AI alone. Be skeptical of specific numbers, dates, and citations — cross-check them against trusted primary sources.

Ask Follow-Up Questions

If something in the AI's response sounds off or feels too convenient, don't let it pass — ask follow-up questions to probe the claim. Hallucinations often unravel under targeted questioning.

Install

Option A: CLAUDE.md (Recommended)

Because anti-hallucination is a guardrail that should apply to every factual claim, CLAUDE.md is preferred — it's always loaded into context, so there's no risk of the skill failing to trigger at the moment you need it most.

A1: Global — one install, all projects (Recommended)

Append to your user-level ~/.claude/CLAUDE.md, which Claude Code loads for every project automatically:

mkdir -p ~/.claude
echo "" >> ~/.claude/CLAUDE.md
curl https://raw.githubusercontent.com/instantX-research/anthropic-anti-hallucinate-skills/main/CLAUDE.md >> ~/.claude/CLAUDE.md

A2: Per-project — share with your team via git

Install into the project root so teammates pulling the repo pick up the same guardrails automatically.

New project:

curl -o CLAUDE.md https://raw.githubusercontent.com/instantX-research/anthropic-anti-hallucinate-skills/main/CLAUDE.md

Existing project (append):

echo "" >> CLAUDE.md
curl https://raw.githubusercontent.com/instantX-research/anthropic-anti-hallucinate-skills/main/CLAUDE.md >> CLAUDE.md

Option B: Claude Code Plugin

From within Claude Code, first add the marketplace:

/plugin marketplace add instantX-research/anthropic-anti-hallucinate-skills

Then install the plugin:

/plugin install anti-hallucinate@anthropic-anti-hallucinate-skills

And reload plugins to apply:

/reload-plugins

This installs the guidelines as a Claude Code plugin, making the skill available across all your projects. Note that skills are loaded on-demand based on the skill's description — which means there's a chance Claude won't recognize the current context as high-risk and skip loading it.

File Structure

.
├── CLAUDE.md                              # Core guidelines (drop into any project)
├── EXAMPLES.md                            # Good vs bad examples for each principle
├── skills/
│   └── anti-hallucinate/
│       └── SKILL.md                       # Claude Code skill definition
├── LICENSE
└── README.md

Customization

These guidelines are a starting point. You can:

• Add domain-specific rules — e.g., "Never guess medication dosages" for medical projects
• Adjust strictness — tighten for research work, relax for creative brainstorming
• Combine with other skills — these guidelines complement coding-style skills like andrej-karpathy-skills

How to Judge Effectiveness

After applying these guidelines, Claude should:

• Say "I don't know" or "I'm not sure" more often (this is a feature, not a bug)
• Provide fewer fabricated citations and statistics
• Clearly distinguish facts from inferences
• Proactively suggest verification steps for high-risk claims
• Self-correct when challenged rather than doubling down

Acknowledgments

• Source: Why do AI models hallucinate? — Anthropic's official Claude channel
• Reference: forrestchang/andrej-karpathy-skills — project structure and formatting style inspired by this repo

License

MIT