foundations-grounding-communication

Installation
SKILL.md

Grounding & Communication Foundations


Scope note — two senses of "grounding": This skill covers the Clark conversational sense: establishing shared meaning between parties (common ground, acceptance evidence, repair). It does not cover the LLM-attribution / RAG sense: whether a generated response is grounded in retrieved documents (citation faithfulness, hallucination detection). For the attribution sense, see ai-rag and the FACTS Grounding / RAGAS / Wallat et al. frameworks (sources FACTSGrounding2025, RAGAS2024, WallatFaithfulness2024 in data/sources.json).


Use the ten primitives below to establish enough shared understanding for the task. Delivered context, an acknowledgment, and demonstrated understanding are different evidence; none grants authority.

Static vs. dynamic grounding. Interpreting an opening brief is different from maintaining a joint plan across turns. Yao, Zou, and Hawkins (2026, preprint) observe dyads missing optimal allocations despite solving the negotiation task individually. Their observed failures motivate preserving settled state, identifying referents explicitly, and distinguishing exploratory proposals from commitments. These are application conventions, not a guarantee that an agent understood.

Laban et al. (2025) measure multi-turn degradation: LLMs lose an average of 39% across six generation tasks in underspecified multi-turn settings, and once a model takes a wrong turn it "gets lost" and does not recover — a reliability failure, not mainly a capability one. Operational inference: in multi-turn underspecified work, periodically consolidate accumulated requirements into one restated spec (#1, #3) rather than patching state in place; after a detected wrong turn, restart execution from the consolidated spec instead of continuing to build on the compromised thread.

When to Apply

Installs
17
GitHub Stars
102
First Seen
Aug 12, 2026