coral

my-project/
+ .claude/
+   CLAUDE.md                 ← project hub: build commands, workflow, critical rules
+   agents/
+     code-critic.md          ← code quality review
+     ...                     ← domain agents (React, Go, ML, infra, etc.)
+   rules/
+     conventions.md          ← naming, git, style
+     ...                     ← domain rules, auto-activated by file path
+ docs/
+   architecture.md           ← module map, dependency graph

Browse this repository's .claude/ folder for a real example.

Each stage produces an artifact that feeds the next. Enter at any point — skip stages you don't need.

pathfind — "I have problems but don't know what to build." Clusters symptoms, investigates root causes (spawns scanner for codebase analysis), generates divergent directions through orthogonal lanes, and spawns pioneer for elegant alternatives. Outputs a ranked direction list with scoring matrix. Hands off the chosen direction to preplan.

preplan — "I know the direction but need agreement on scope." Fills a 7-item agreement (problem statement, success criteria, scope, assumptions, affected systems, constraints, approach direction) autonomously from codebase analysis, then presents to the user for correction. Spawns pioneer to find elegant alternatives for uncertain items, offering default/minimal/elegant spectrums. Produces pre-{topic}.md — the contract that plan must satisfy.

plan — "I need a design before I build." Multi-round review loop: dispatches architect + critic + resolver as a workflow, synthesizes findings, edits the plan file, and evaluates an exit gate (round=N sets the round budget). The resolver classifies findings (Adopt/Adapt/Defer/Diverge), applies changes, and decides whether another round is needed based on finding severity and nature. Produces {topic}.md with acceptance criteria, implementation phases, and execution order (dependency graph + parallel batches).

ralph — "I have a plan (or prompt). Just build it." Persistent executor with verification loop. In plan mode, reads the execution order and dispatches batches — parallelizing independent ACs. Every completion claim requires fresh verification evidence (lint → build → test). --red spawns a red-attacker in parallel to write adversarial tests targeting blind spots. --team uses Agent Teams for parallel AC execution.

# Deeper iteration — up to 3 review rounds per phase:
/coral:plan round=3 add retry logic to the API client

# Adversarial testing — spawns a red-attacker to target blind spots:
/coral:ralph --red implement the caching layer