A deliberation facilitator for Claude Code. Named after the human in the movie, not the machine.
Dave is a skill that guides structured thinking on open questions across sessions. It sits in the known-unknowns quadrant: questions you know you're carrying but can't yet resolve. Evidence accumulating from different contexts. Tensions forming between candidate positions. The question itself still shifting shape.
Most AI tools want to answer your question or execute your task. Dave does neither. It holds the process while you do the thinking.
Dave facilitates deliberation through epistemic moves: some widen the space, some compress it toward resolution, some reframe the question entirely. It tracks questions, evidence, tensions, and positions in a persistent event log (.dave/questions.jsonl), so you can pick up a question after weeks and see exactly where you left off.
One rule: the human is the cognitive agent. Dave holds methodological structure. It never resolves the question for you.
skills/dave/
SKILL.md # Core facilitator process and iron law
METHODS.md # Epistemic moves, lifecycle shifts, calibration signals
EXPLORER.md # Subagent guidance for deep material exploration
STATE.md # Event schema and field definitions
hooks/
deliberation-context.sh # Session-start hook, surfaces open questions
docs/
philosophy/ # Core principles, design reasoning
research/ # Research landscape and references
plans/ # Design and implementation plans
Copy the skills/dave/ directory into your Claude Code skills path and the hook into your hooks configuration. See the Claude Code docs for skill and hook setup.
The design draws on five research threads:
- Linus Lee on instrumental vs. engaged interfaces: deliberation is engaged, the process is the value (summary, interview)
- Collaborative Causal Sensemaking (arxiv): agents should track the human's evolving world model, not resolve uncertainty on their behalf
- Giuseppe Riva on cognitive scaffolding: frictionless AI induces complacency, thinking tools need calibrated challenge
- Ought/Elicit on process supervision: supervise reasoning processes, not outcomes (plan)
- GESIS/Nokia Bell Labs on agent-assisted foresight: agents expand the search space, humans own the sensemaking
Full research notes in docs/research/landscape.md.
Working prototype. Built for one user, one practice. Not packaged, not polished, not generalized.
MIT