A conversational specification tool. An agent interviews you about your project and creates complete specs before any code is written.
-
Spec Loop Model - Spec creation is a conversation, not a loop. A human and Claude Code sit down, discuss, and produce specs. Learning is the bigger loop — it wraps around building (spec → build → learn → spec).
-
Assumption Validation - Surface assumptions explicitly, categorize by area (Desirability/Feasibility/Viability), rate importance 1-10, and test what can be tested. High-importance assumptions with no evidence are blind spots that gate progress.
-
Learning Loop - Specs aren't static. Build → observe → learn → update specs. The cycle closes when evidence flows back into spec updates.
-
Trust Model - Trust starts low and builds through good outcomes. Agent earns autonomy over time, from "approve everything" to "YOLO mode." Personalized to individual preferences.
-
Spec Observability - Tracking layer (database-backed) maintains assumptions, experiments, evidence, and learnings as structured entities separate from specs. Soft gating: agent flags high-importance unchecked assumptions and waits for human decision.
-
Spec Lifecycle - Two states: draft ⇄ ready. Human decides transitions. The spec review agent annotates.
Start Claude Code and run the spec agent:
/agents/spec.sh
Or paste a prompt manually — study the specs, then interview.
- specs/ - Design documentation
- agents/ - Agent scripts and prompts
- the-ralph-playbook.md - Source material, the workflow this is based on