A Claude Code skill: two small gates to run before a confident or novel claim ships.
Reasoning that never leaves the abstract, and reasoning that never touches anything outside itself, both fail in predictable ways. The first produces claims too general to be wrong. The second produces claims that sound settled but were never actually checked against anything. Neither failure is visible from the inside — a closed loop of reasoning looks perfectly consistent to the loop itself.
grounding-check is two questions to ask before a claim is presented as finished:
- Altitude — is this pitched at a useful level between one concrete instance and a general principle, or is it stuck at one end?
- External reference — does this touch something that didn't come from your own reasoning in this context (a search, a file you actually opened, a question you actually asked), or is it entirely self-referential?
See skills/grounding-check/SKILL.md for the full
skill.
The idea of checking AI output against something outside the model's own reasoning isn't new — it's an established pattern, sometimes called "citation grounding." grounding-inspector (confirmed by reading its README) decomposes AI output into atomic claims and verifies each one against source document chunks after the fact. A search also surfaced a listing for "Environment Grounding Automation," described as checking a plan against live system state before acting — but its page returned a bot-check wall on every fetch attempt, so that description is unverified and should be weighted accordingly. What this skill adds is narrower: a pre-emission self-check paired with a second, independent axis (concrete/abstract altitude) — meant to run on the reasoning itself, before a claim is written down, rather than verifying already-written output against a document or a system.
Copy the skill folder into your project's skills directory:
cp -r skills/grounding-check .claude/skills/or into ~/.claude/skills/ to make it available across projects.
MIT — see LICENSE.