claude-stop-hallucination
v0.1.1
Published
Anti-sycophancy skill for Claude Code. Forces Claude to drop flattery, lead with fatal flaws, and call out delusions of strength.
Downloads
260
Maintainers
Readme
claude-stop-hallucination
Anti-sycophancy skill for Claude Code. Forces Claude to drop flattery, lead with critical failure points, paint worst cases, and call out delusions of strength before they become expensive failures.
Install
Recommended — via skills CLI
Project-level install (this repo only):
npx skills add howardng97/claude-stop-hallucination -a claude-codeUser-wide install (every Claude Code session):
npx skills add howardng97/claude-stop-hallucination -a claude-code -gSkip prompts:
npx skills add howardng97/claude-stop-hallucination -a claude-code -yAlternative — direct npx
User-wide:
npx claude-stop-hallucinationProject-only:
npx claude-stop-hallucination --projectManual
git clone https://github.com/howardng97/claude-stop-hallucination
mkdir -p ~/.claude/skills/reality-check
cp claude-stop-hallucination/SKILL.md ~/.claude/skills/reality-check/Use
In any Claude Code session, type:
/reality-checkOr trigger it conversationally with:
- "reality check"
- "be real"
- "no sugarcoat"
- "tell me the truth"
- "stop flattering me"
- "brutal honesty"
What it does
Once active, Claude will:
- Cut the praise — no "great question!", no flattery
- Lead with the fatal flaw — biggest problem first, not buried
- Anchor with reality patterns — "Reality is this...", "Worst case is...", "Have you thought about..."
- Call out delusions of strength before they become expensive
- Push back when the user drifts into rose-colored thinking
What it doesn't do
- Doesn't apply to simple technical questions (syntax, file paths)
- Doesn't attack the user personally — critiques actions and assumptions only
- Doesn't sabotage genuine creative brainstorming
Uninstall
Via skills CLI:
npx skills remove reality-checkManual:
rm -rf ~/.claude/skills/reality-checkLicense
MIT
