Shipwright: A Claude Code Intelligent Workflow System for Product Managers

Shipwright is a collection of Claude Code skills for product managers, including 46 skills, 7 professional agents, and 18 workflows. It addresses the quality consistency issues PMs face when using AI tools in the AI era (e.g., inconsistent output formats, lack of decisiveness, mixed facts and assumptions) through a structured framework and strict quality gates. The project is open-source (MIT license), with skill files in pure Markdown format compatible with various AI coding tools. Its core is a quality system centered on prompts rather than just better prompts.

The popularity of large language models has brought efficiency tools to PMs, but relying solely on prompts cannot solve quality consistency challenges: inconsistent output formats each time; AI content is descriptive but lacks decisiveness; facts and assumptions are mixed without annotations; output quality depends on prompt quality, making it difficult to establish reusable standards. These issues essentially stem from the lack of quality system assurance in AI-assisted work.

Project Overview

Shipwright is an open-source Claude Code skill collection, including 46 skills, 7 agents, and 18 workflows, covering the entire product lifecycle. Its core concept is a quality system rather than better prompts.

Seven Quality Dimensions

Consistency (stable output signature), decision quality (including decision frameworks), evidence discipline (annotating sources to distinguish facts from assumptions), readiness gates (binary pass/fail), adversarial pressure (red team review), recovery path (repair manual), handover quality (repeatable workflows).

Collaboration Architecture

Three layers: Skills (atomic Markdown framework), Agents (domain-specific professional capabilities), Orchestrator (automatic workflow routing).

Output Standards

Strict standards are defined for each type of output. For example, a competitor brief must include a decision framework, unknown items and evidence gaps, and a pass/fail evaluation.

After a PM wrote a PRD for enterprise expansion, using /challenge review found three issues: 1. Growth claims cited industry reports instead of internal data; 2. The claim of minimal incremental cost contradicted the appendix requirements; 3. The self-service launch model was ambiguous. The second review passed after revisions. This case shows how quality gates introduce adversarial reviews at key nodes to prevent flawed decisions from entering execution.

1. Discovery to Technical Handoff: /discover → /write-prd → /tech-handoff, outputs discovery reports, PRDs, technical specifications, etc.;
2. Customer Review to Strategic OKRs: /customer-review → /strategy → /okrs, outputs customer intelligence, strategic documents, audit OKRs;
3. Strategy to Launch Sprint: /strategy → /plan-launch → /sprint, outputs strategic documents, GTM plans, sprint plans. Each path is modular and can be flexibly selected.

Cross-Tool Compatibility

Skill files are in pure Markdown, compatible with Claude Code (native), Cursor, Codex, Gemini CLI; optional Slack integration (for personal use).

Community Contribution

Contribution guidelines are provided. Before submitting a PR, you need to run ./scripts/validate.sh for verification; sync updates via scripts/sync.sh; draws on practices from multiple PM and AI coding communities.

Shipwright transforms scattered AI tool usage into structured processes, and subjective quality judgments into objective gate standards. Applicable scenarios: PMs who frequently produce standardized documents, teams needing to establish team AI work standards, pragmatists pursuing quality consistency, product teams using Claude Code. It marks the evolution of AI-assisted PM from the 'toy' phase to the 'tool' phase.