Review and validate project assumptions with Common Ground
Run Common Ground in Claude Code to surface project assumptions, confirm their evidence and confidence with the operator, and save a reviewable Markdown and JSON record before implementation.
npx skills add agentskillexchange/skills --skill review-and-validate-project-assumptions-with-common-ground
Invoke Common Ground before a feature, unfamiliar-codebase investigation, or major refactor when the agent’s assumptions about architecture, testing, coding standards, or user preferences need explicit review. This workflow uses the Common Ground command distributed in Jeffallan’s Claude Skills plugin for Claude Code.
What this skill actually does
Install the upstream marketplace with /plugin marketplace add jeffallan/claude-skills, then /plugin install fullstack-dev-skills@fullstack-dev-skills, and restart Claude Code when prompted. Use the plugin installation path because the skills-only installation methods omit slash commands. In the intended project, run /common-ground. The command identifies the project from its Git remote, falling back to its absolute path, and inspects configuration, conversation context, and any existing grounding record.
The agent presents assumptions for operator selection, records whether each was stated, inferred, assumed, or uncertain, and proposes ESTABLISHED, WORKING, or OPEN confidence tiers. The operator confirms corrections and tier changes before the record is written. Outputs are COMMON-GROUND.md and ground.index.json under the project’s ~/.claude/common-ground/ directory, with sources, timestamps, and change history.
Use /common-ground --list for a read-only view and /common-ground --check to revalidate assumptions after requirements change. The optional --graph mode produces a reviewable diagram of the reported decision structure after assumptions are confirmed. This scope ends at assumption review and record maintenance; implementing code, running other bundled project commands, and connecting Jira or Confluence are separate tasks. Confidence labels and diagrams are agent-authored aids for human review, not proof of correctness or access to hidden model internals.