Designed for Claude Code. Not just AI-compatible, but AI-directed. A complete methodology for building with Claude Code, from planning to deployment. All documentation is structured for token-efficient AI consumption, with progressive disclosure that loads only what’s needed.
Skills are markdown-based capabilities that teach Claude Code how to handle specific tasks. They cover agent orchestration, code review, database migrations, design system enforcement, email integration, frontend patterns, git workflows, landing page generation, security scanning, test scaffolding, and more. Each skill includes context, constraints, and patterns for its domain. 15 of them ship with regression-tested evals — automated pass/fail checks against a baseline scorecard — so a change to a skill can’t silently degrade its quality.
Sub-agents handle complex tasks autonomously: bug fixing, bundle analysis, changelog generation, database migrations, documentation writing, migration review, test healing, issue creation, plan review, code refactoring analysis, deep research, security scanning, spec planning, and test writing. Agents have persistent memory, timeout control, and permission modes (read-only, file editing, or fully autonomous).
Hooks run automatically during Claude Code sessions. They block destructive git operations, validate database migrations before execution, inject project context (branch, issue info) into every prompt, scan for prompt injection attempts, audit security-sensitive commands, enforce workflow rules, and guard worktree file isolation. Two-layer permission model: settings plus hooks.
The Docker MCP Gateway connects Claude Code to the six services your app runs on, so Claude doesn't just write code for them, it operates them from chat. Ask for yesterday's signups or a funnel drop-off and Claude queries PostHog, including AI token usage and cost. It checks Vercel build and runtime logs when chasing a bad deploy, runs Supabase migrations when shipping a schema change, and refunds a customer through Stripe. Email and cache (Resend, Upstash) round out the set. This closes the loop: the agent that finds a problem in your telemetry can fix it and ship the change, without you copying data between dashboards.
Simple workflow for low-risk changes (80% of work): /start → /code → /commit. Spec-driven workflow for complex features: /start → /spec → /plan → /tasks → /code → /verify → /commit → /pr. The /start command auto-detects which lane to use based on the issue’s complexity and risk domain.
The slash commands you'll reach for most, organized by what you're trying to do.
/thinkExplore a feature idea before writing code/bootstrapTurn a PRD into a GitHub project board/startBegin work on a GitHub issue/commitCreate a formatted conventional commit/prCreate or update a pull request/cleanupRemove merged branches after a PR lands/testGenerate tests for a new feature/designCheck a new page or component against the design system/learnCapture a lesson from what just happened/reflectAnalyze lessons and suggest codebase improvements/healthAudit architecture after heavy development/updatePull latest Sinter updates into your project/setupWalk through initial project configuration/dbDatabase schema changes and migrationsEvery layer is designed to minimize token consumption. Skills load in three tiers: metadata first (~50 tokens), usage details on demand (~500), full guides only when needed (~2,000). Code search uses AST patterns instead of reading full files, reducing research from ~30K to 3-5K tokens. A .claudeignore file keeps archives and build output out of context. Session hooks inject branch and issue state once at startup. Heavy operations run in isolated subagent contexts. Workflows checkpoint to disk for near-zero-cost resumption. On a worked example, an automation task ran with about 89% fewer tokens (around 52% on LLM-assisted tasks) compared to an unoptimized codebase.
$249 one-time. Lifetime access. Unlimited projects.
Get Sinter: $249