zaksak

A collection of Claude Code utilities by Zak Sakata.

A spec-driven development workflow that turns a rough idea into a verified implementation. It elicits and hardens an acceptance-criteria-first spec against your codebase, compiles it into an execution driver, then checks the built code against every criterion — with read-only adversarial judges gating quality at each stage and every spec change committed automatically. When OpenAI Codex CLI is available it adds an optional, fail-open second judge of a different provider, so every quality gate is cross-checked rather than one model auditing itself.

Ralph Loop — an autonomous agentic coding workflow with sandbox isolation, scoped memory, and self-managing task iteration.

Manage your Claude Code skills and plugins from inside any project — switch which plugins are active, publish and sync local skills to a central marketplace, and diagnose and auto-repair a broken setup. Adapts to any repo layout and delegates skill authoring to the native skill-creator.

A git worktree workflow for running multiple Claude Code sessions in parallel — no restarts, no manual checkouts. Built on native worktrees, it covers the full lifecycle from spinning up an isolated worktree to finishing it through a guarded, never-squashed PR merge, with a live dashboard of in-flight work and deterministic guards against squash merges and committing unresolved conflicts.

Run a small team's entire software lifecycle on GitHub Projects v2 — deterministic, free, GitHub-native, and with no metered AI. It provisions a golden-template board and repo automation across an org, then drives work from issue intake through sprint planning and the full branch → PR → merge flow with check-gated, never-squashed merges, kept safe by GitHub App-token writes and skill-scoped guards.

Run the OpenAI Codex CLI from inside Claude Code as a cross-model escape hatch: hand a one-off review, question, or write task to a different provider's model for a second opinion, grounded in your repo with web search always on. Every call routes through one fail-open bridge — if Codex is absent, unauthenticated, or broken the skill reports and stops rather than answering as Claude, and Codex's output is surfaced verbatim as untrusted external text. Assumes an OpenAI-authenticated Codex setup.