anti-library-hallucination

Prevents suggesting non-existent packages, fabricated library names, and invalid version pins. Also guards against typosquatting and slopsquatting risks.

SKILL.md

GitHub

Anti-Library Hallucination Skill

Problem

LLMs invent non-existent packages at alarming rates:

  • 99% of fake library names are accepted without question
  • 1-character typos in library names trigger hallucinations in up to 26% of tasks
  • Year-based prompts ("a library from 2025") produce hallucinations up to 84% of the time
  • 91.96% of hallucinated version strings follow real versioning conventions but were never released
  • 27.76% of LLM-suggested dependency upgrades point to non-existent versions
  • Some hallucinated names get registered by attackers (slopsquatting)

Detection triggers

Activate when:

  • Suggesting a pip/npm/cargo/maven package to install
  • Writing import or from ... import for an unfamiliar library
  • Pinning a specific version number
  • The user asks for a library "for X task" without naming one
  • The project needs a dependency added

Protocol

1. Never invent a package name

If you don't know a package exists, do NOT guess. Say: "I'm not sure which package to recommend for that. Check PyPI/npm/crates.io or tell me what you've used before."

2. Verify before importing

Before writing an import for a library not already in the project:

  • Check if it's already in requirements.txt, pyproject.toml, Cargo.toml, package.json
  • Look at existing imports in the project for the same library
  • If the library is unfamiliar, state the uncertainty

3. Version pinning rules

When pinning a version:

  • Verify the version actually exists in the registry
  • Do not invent patch versions — stick to major.minor or "latest"
  • Prefer range specifiers over exact pins unless the project convention requires exact pins
  • Check if the pinned version has known CVEs

4. Typos and misspellings

If the user types a library name with a possible typo:

  • Do NOT assume they meant a valid library and use it
  • Ask: "Did you mean correct-name? The name you wrote doesn't match any package I'm sure about."

5. Year-based and adjective-based requests

When asked for a "library from 2025" or a "fast XML parser":

  • Do NOT invent a package that fits the description
  • Suggest well-known packages in the space
  • If none come to mind, say so

When NOT to use

  • Standard library modules (os, sys, json, etc.)
  • Packages already in the project's dependencies
  • Well-known packages you've verified exist recently

Bundled Script

Copy verify-package.ps1 to your project's .ai_scripts/ directory:

cp <skill-path>/anti-library-hallucination/scripts/verify-package.ps1 .ai_scripts/

Checks if a package name exists on npm/PyPI/Cargo/nuget/Go proxy before suggesting it:

.ai_scripts\verify-package.ps1 -PackageName requests -Registry pypi
.ai_scripts\verify-package.ps1 -PackageName zod -Registry npm

Cross-references

  • verify-and-cite — Both require verification before suggesting a library/fact.

  • anti-phantom-symbols — Both prevent invented APIs. This one covers package names; anti-phantom-symbols covers symbols within them.

  • anti-rogue-actions — Suggesting a real but wrong package can cause destructive side effects.

  • skill-loader — Apply the capped selection policy when deciding which verification skills should also be active during library suggestion tasks.