The Two Phases

Everything Leopold does splits into two phases: a brief you co-author, and a run it conducts.

stateDiagram-v2
    [*] --> Brief
    Brief --> Brief: debate / revise
    Brief --> Run: /leopold-run
    Run --> Run: turn → decide → next item
    Run --> Escalated: irreversible + ambiguous
    Escalated --> Run: you decide, re-run
    Run --> [*]: plan complete

Phase 1 — Brief

A structured, adversarial conversation that ends with four durable artifacts in .leopold/:

Artifact	Answers	Role
MISSION.md	what are we building, and what is "done"	the goal
CHARTER.md	how would you decide	the decider — "becomes you"
GUARDRAILS.md	what stays locked	the safety boundary
PLAN.md	in what order	the work queue

The brief is the contract. The run never invents intent; it executes the brief. Because the run is capped by the quality of the brief, this phase is where your attention pays off most.

Phase 2 — Run

Leopold reads the brief and conducts Claude Code through the plan. It decides forks from the charter, logs every non-mechanical call, and only stops for the rare fork that is both irreversible and ambiguous.

flowchart LR
    Plan[(PLAN.md)] --> N[next item]
    N --> W[work it]
    W --> Fork{fork?}
    Fork -- no --> Mark
    Fork -- "decide from charter" --> Log[(DECISIONS.md)] --> Mark[mark done]
    Mark --> Plan

Two engines implement Phase 2:

• The in-session engine — one Claude Code session, kept going by a stop hook.
• The SDK driver — a persistent conductor that spawns fresh workers per item.

Both obey the same Decision Protocol and Guardrails.