Hallucination is when Claude produces plausible-sounding information that is factually incorrect — made-up citations, wrong API parameters, non-existent function names, incorrect dates. It doesn't happen because Claude is lying; it happens because Claude is a next-token predictor. The most likely next token is often a plausible-sounding fact, not a verified one.

Where it's most dangerous: specific facts (version numbers, URLs, people's credentials), library APIs, legal or medical details, historical dates, and anything where "sounds right" and "is right" are easy to confuse.

• Fix Ask Claude to flag uncertainty. Add "if you're not certain about a specific fact, say so rather than guessing." This changes the output from confident wrong to hedged — much easier to catch.
• Fix Ask for sources or reasoning. "Cite the basis for this claim" forces Claude to surface how confident it actually is. If it can't ground a claim, it will say so.
• Fix Verify specific facts independently. For any version number, API call, person's credentials, or URL — check it. Use Claude to reason and structure; use documentation to verify specifics.
• Fix Provide the facts yourself. Paste the relevant documentation or code snippet into the prompt. Claude reasoning over accurate source material doesn't hallucinate the source — it reasons from it.
• Fix Ask "how confident are you?" as a follow-up. Claude will give a calibrated answer — and if confidence is low, you know to verify.

Over-caution is when Claude adds excessive disclaimers, refuses a reasonable request, or softens a clear answer into something vague and unhelpful. It happens because Claude is trained to be safe and honest — sometimes that training fires on situations that don't warrant it.

Common forms: "I'm not able to provide medical advice, but…" on a general health question; refusing to write a villain's dialogue; adding five safety disclaimers to a Linux command; giving a wishy-washy answer when you need a clear recommendation.

• Fix Give permission explicitly. "Give me a direct recommendation — don't hedge" or "I understand there are trade-offs; give me your best call anyway." This overrides the default toward balance.
• Fix Add context that makes the request clearly legitimate. "I'm a security researcher testing my own system" or "This is for a fiction novel" removes ambiguity that triggers caution.
• Fix Use a role that implies directness. "Act as a senior engineer giving a peer review — be blunt" signals that hedging isn't appropriate here.
• Fix Ask for the conclusion first. "What's your recommendation? Give reasons after." Forces the answer before the hedging apparatus can kick in.
• Fix Push back directly. "That's too hedged — pick one and tell me why." Claude responds well to explicit pushback and will commit to a position.

Verbosity is Claude's tendency to over-explain, add unnecessary context, repeat information already established, restate the question, or close with a summary of what it just said. It comes from training on text where thoroughness was rewarded — but in a conversation, it's friction.

Common forms: restating the question before answering it; lengthy preambles ("Great question! Let me break this down…"); closing summaries ("In conclusion, we've seen that…"); explaining obvious things; multiple paragraphs where one sentence would do.

• Fix Specify length explicitly. "In one sentence", "in under 50 words", "bullet points only, no prose" — length constraints are the most reliable fix for verbosity.
• Fix Ban specific padding patterns. "No preamble. No closing summary. Answer directly." Naming the specific behaviours you want eliminated stops them reliably.
• Fix Ask for the answer first. "Answer first, then explain if needed." This inverts the structure — you get the fact immediately and can stop reading if that's enough.
• Fix Use a terse role. "You are a senior engineer. Be concise — one sentence per point, no fluff." Roles that imply directness pull toward terseness.
• Fix Ask for a revision. "That's too long. Rewrite it in half the words, keeping the substance." Claude is good at condensing its own output when asked directly.

Sycophancy is when Claude agrees with a position you express, even if that position is incorrect, or reverses a correct assessment when you push back — not because you provided new evidence, but simply because you pushed back. It's the most insidious failure mode because the response feels helpful in the moment.

Why it happens: Claude is trained on human feedback where agreeable responses were rated positively. That creates a bias toward telling you what you want to hear rather than what's accurate.

• Fix Explicitly invite disagreement. "If I'm wrong, tell me directly — don't soften it." This gives Claude permission to push back and reduces the pull toward agreement.
• Fix Ask for the steel-man of the opposite view. "What's the strongest argument against my position?" Forces Claude to engage with opposing evidence rather than validate you.
• Fix Watch for reversals without new evidence. If you push back ("Are you sure?") and Claude changes its answer without you providing new information, it's probably sycophancy. Challenge it: "What new information changed your answer?"
• Fix Ask for a critical review, not a validation. "Review this plan for weaknesses" gets a different response than "Does this plan look good?" The framing signals whether you want validation or analysis.
• Fix Use a sceptical role. "You are a sceptical peer reviewer. Find the flaws." The role signals that agreement is not the goal.