claude-plugins

A Claude plugin marketplace covering two halves of the same job: turning an idea into shipped, tracked work, and turning what you learn along the way into notes you can find again. Both ship as composable, natural-language-triggered skills that run in Claude Code and Cowork.

Two plugins, thirteen skills.

Plugin	Skills	For
delivery-design	9	An end-to-end SDLC pipeline: idea → PRD → tech spec → delivery plan → tickets → Linear/GitHub
knowledge	4	Note and document authoring for an Obsidian vault: what shape a document should take, and the workflows that produce it

They install independently and share no files. Install one, the other, or both.

Design principles

Both plugins are built around a small set of deliberate constraints rather than an accumulation of features:

• One source of truth: local markdown. Project-management tools (Linear, GitHub Issues) are sync targets, never the origin. State lives where the code lives.
• A document is worth writing only if its "why" can't be recovered from the code. That single test gates every decision record — it keeps documentation from drifting into noise.
• Don't reinvent formats that already have standards. Decisions use MADR-style records; agent context uses AGENTS.md; nothing bespoke where a convention already exists.
• Formats are data; skills are workflows. A shape is defined once, in the skill that owns it. Every other skill owns a process and delegates shape. If a convention appears in two skills, that is a bug.
• Skills never hardcode where a file goes. Placement, naming, and folder structure are read from the host project or vault at runtime, so the same skill works in any of them.

delivery-design

PRODUCT workflow  (new product)
  product-brief → domain-doc → product-backlog

FEATURE workflow  (per backlog item)
  feature-request → feature-brief → feature-design → feature-plan → feature-tickets → feature-publish
      intake          idea→PRD        PRD→spec         spec→plan       plan→files       files→Linear/GitHub

decision-record is available throughout, at any point a costly, hard-to-reverse decision is made.

A brand-new product runs the product workflow first, then pushes each backlog item through the feature workflow. A new feature on an existing product skips straight to feature-brief.

Skill	Does
product-brief	Idea → Product Brief + MoSCoW-prioritised backlog
domain-doc	Bootstraps and maintains the glossary + domain model
feature-request	Intake triage — the front door before a PRD
feature-brief	Feature idea → PRD
feature-design	PRD → Technical Specification (runs a virtual dev team of specialist lenses)
feature-plan	PRD + Tech Spec → sequenced Delivery Plan
feature-tickets	Delivery Plan → local Epic + Story ticket files
feature-publish	Local ticket files → Linear or GitHub Issues
decision-record	MADR-style decision record with a significance gate that refuses trivial ones

knowledge

note-formats is the source of truth for document shape and house style. The other three own workflows and call it rather than restating it.

Skill	Does
note-formats	Decides which of eleven shapes a document should take, and defines each one
obsidian-note-writer	Create, enhance, and refactor notes — search before creating, safe splits and merges
session-synthesis	Conversation or transcript → conceptual map → notes
ops-docs	A messy process → Checklist, Runbook, or Playbook

The eleven shapes split into knowledge shapes, read to build understanding (Atomic Note, Overview, Hub, Decision Guide, Guide, Map of Content), and operational shapes, read to get something done (Checklist, Runbook, Playbook, Standard Note, Decision Record). Choosing between them is the hard part and the reason the skill exists — references/choosing-a-format.md carries the criteria, the pairs that get confused, and worked examples.

Repository structure

.claude-plugin/marketplace.json      marketplace catalog (fixed path)
plugins/
├── delivery-design/
│   ├── .claude-plugin/plugin.json   plugin manifest (versioned)
│   └── skills/                      one directory per skill: SKILL.md + templates + subagent prompts
└── knowledge/
    ├── .claude-plugin/plugin.json
    └── skills/
        └── note-formats/
            ├── references/          doctrine: choosing a format, structure, Obsidian style
            └── assets/              one skeleton per shape

Skills are auto-discovered by walking skills/*/SKILL.md; they are not declared in the manifest.

Install

/plugin marketplace add edwinwright/claude-plugins
/plugin install delivery-design
/plugin install knowledge

Skills then trigger on natural language — see each skill's SKILL.md for its trigger phrases.

KNOWN-ISSUES.md documents an upstream Cowork bug that can leave installed plugins showing a stale version of this repo. If a skill behaves as though a recent change never happened, start there.

Status

All thirteen skills ship and are in active use. CLAUDE.md carries the conventions these plugins are developed under, including why two overlapping skills were merged into one.

Licence

MIT — see LICENSE.