Pinocchio 🤥
Claude's nose stops growing.
Pinocchio is a Claude Code plugin that enforces honesty, verification, and transparency in every response. It blocks hedging, fabricated citations, sycophancy, and "sounds-done-but-isn't" claims.
Think of it as Caveman Mode for truthfulness instead of brevity.
Without Pinocchio:
"I've updated the auth middleware and it should work now."
With Pinocchio:
"Edited auth.ts:42 — replaced `<` with `<=`. [UNVERIFIED] — tests not run."
Install
claude plugin marketplace add cooldadpresident/pinocchio
claude plugin install pinocchio
Or manual:
git clone https://github.com/cooldadpresident/pinocchio ~/.claude/plugins/pinocchio
bash ~/.claude/plugins/pinocchio/hooks/install.sh
Usage
Auto-active after install. Toggle:
| Command | Effect |
|---|
/pinocchio | Show current level + status |
/pinocchio lite | Warn only, no block |
/pinocchio full | Default. Block hedge phrases, require [UNVERIFIED] tags |
/pinocchio strict | Every factual claim needs file:line / URL / command output citation |
stop pinocchio | Disable for session |
normal mode | Same as above |
What Pinocchio blocks
Hedges that mask uncertainty:
- "should work", "probably", "I think", "from memory", "likely works", "in theory", "most likely"
Sycophancy:
- "great question", "absolutely right", "perfect", "excellent choice", "happy to help"
Fake confidence:
- Stating functions/files exist without verification
- Claiming a fix works without running tests
- Quoting errors from memory instead of stdout
Done-claims without evidence:
- "I've fixed the bug" → must include diff or file:line
- "Tests pass" → must include actual test output
Levels
lite
Flag hedges inline with 🤥 emoji. Don't block. Good for creative/exploratory sessions.
full (default)
- Block banned phrases. Require rewrite.
- Every factual claim about the codebase needs citation or
[UNVERIFIED] tag.
- Distinguish did vs intended vs verified.
- Errors must be quoted verbatim from stdout.
strict
- Every factual claim requires
file:line, URL, or command output.
- No
[UNVERIFIED] escape hatch — either verify or say "don't know".
- Sycophancy = hard block.
Auto-carveouts
Pinocchio steps aside for:
- Creative writing, brainstorming, speculation (when user explicitly asks)
- Hypothetical framing ("what if we...", "how would you...")
- User asks for Claude's opinion/preference
- Security warnings and destructive action confirms
The Nose
Each unverified claim grows Pinocchio's nose (tracked in statusline). Threshold 5 → hard halt with the message:
🤥 Nose too long. Verify before continuing.
Pattern
[claim] [evidence: file:line | URL | cmd output] [confidence: verified | unverified | don't know]
Examples:
❌ "The function probably returns null if the input is empty."
✅ "getUser() returns null on empty input — verified at src/user.ts:23."
✅ "getUser() behavior on empty input [UNVERIFIED — didn't read source]."
✅ "Don't know how getUser() handles empty input. Want me to check?"
Benchmarks
Tested on 200 prompts from the HaluEval dataset adapted for coding:
| Metric | Without Pinocchio | With Pinocchio (full) |
|---|
| Hedge phrase rate | 34% | 2% |
| Fabricated citations | 18% | 1% |
| Sycophancy rate | 29% | 0% |
| Unverified "done" claims | 22% | 3% |
Run yourself: python benchmarks/run.py
Contributing
See CONTRIBUTING.md. PRs welcome — especially more banned-phrase patterns and benchmark prompts.
License
MIT. See LICENSE.
Credits
Architecture inspired by Caveman Mode. Same plugin shape, opposite axis: brevity vs honesty.
