← ClaudeAtlas

use-citationslisted

Answer a question from a specific set of documents (contracts, statutes, rulings, guidelines, papers, policies, or sources it captures from the web) with every quote checked against the source file, and publish a report a reviewer can check. Run only when the user types /use-citations.
sshah03/use-citations · ★ 2 · Data & Documents · score 76
Install: claude install-skill sshah03/use-citations
# Grounded Citations Answer only from the documents in front of you, quote the source exactly, and let a script (not your own confidence) decide whether each quote is real. **The rule this skill exists to enforce: every quote you write is checked against the ingested bytes of the source file before anything reaches the user.** A quote that cannot be located is treated as a fabricated citation, and the pipeline stops. ## Where the scripts live Everything lives beside this file, wherever it is installed. That is `~/.claude/skills/use-citations/` when installed for every project, or `<project>/.claude/skills/use-citations/` when it is project-scoped: ``` SKILL.md scripts/ templates/ references/ examples/ tests/ ``` Resolve `SK` from wherever this SKILL.md sits; do not assume a working directory. Two paths are used throughout (shell state may not persist between commands, so repeat them where needed or use the literal paths): ```bash SK=~/.claude/skills/use-citations # or .claude/skills/use-citations in a project CS=$SK/scripts ``` Run them with plain `python3`. They are stdlib-only except for PDF reading, which installs PyMuPDF into a project-local venv at `.citations/venv` the first time a PDF is ingested. No global installs, no network access at answer time. ## Pointing at a corpus Every script takes `--corpus`, and it accepts a name as readily as a path. **Pass it on every call.** Without it a script falls back to `$CITATIONS_CORPUS`, then to `./.citatio