convez.ai
Skill
AnjanJ

shipkit

AnjanJ/shipkit

Ver no GitHub
InicialPluginScore · de 0 a 100

Ship better code with Claude. Skills, agents, rules, and knowledge bases for Claude Code. Mirror of codeberg.org/AnjanJ/shipkit

Instalação

$npx convez-ai AnjanJ/shipkit

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
3
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 1 dia
Versão
v2.7.0
Issues abertas
1
Contribuidores
2

README

A tradução para o português deste README ainda está na fila.

Shipkit

Buy Me A Coffee Sponsor

Ship better code with Claude. A plugin for Claude Code that gives you workflow rules, skills, agents, knowledge bases, and path-scoped rules — out of the box.

User Guide — detailed docs for every skill, agent, setup/unsetup, and common workflows.  ·  Changelog — what's new.

New in 1.2: the project elders (grandfather, eve, archivist) answer questions about your code without bloating your main context — ask via /shipkit:ask. Optional MemPalace adds decision-history recall. Plus /shipkit:plan for plan-before-code. See the Changelog.

Install

/plugin marketplace add https://codeberg.org/AnjanJ/shipkit.git
/plugin install shipkit@shipkit

Restart Claude Code after installing to load the plugin.

Or test locally:

claude --plugin-dir ~/code/shipkit

That's it. All skills, agents, rules, and knowledge bases are immediately available.

Want to go further? Run /shipkit:setup to tailor everything to your specific project — it auto-detects your stack, creates a CLAUDE.md, and installs stack-specific skills and rules. This is optional but recommended for the best experience.

What You Get Instantly

21 skills — slash commands for common workflows:

SkillWhat It Does
/shipkit:setupConfigure for your stack (Rails, React, Python, Go, Elixir, static)
/shipkit:unsetupRemove setup and restore your project to its pre-shipkit state
/shipkit:askAsk the project elders a question — routed to a subagent, keeps main context thin
/shipkit:mapBuild/refresh a project's PROJECT_MAP.md and register it for cross-project answers
/shipkit:planPlan before coding — PRD, tech spec, and atomic task breakdown (runs in a forked context)
/shipkit:qa5-phase QA workflow with probing questions before writing tests
/shipkit:review-my-code8-lens code review (Clean Code, DRY, KISS, YAGNI, Idioms, Framework, Perf, Errors)
/shipkit:testAuto-detect test framework and run tests
/shipkit:tddTest-driven development — Red-Green-Refactor with Iron Law enforcement
/shipkit:debugSystematic root-cause debugging — investigate before fixing
/shipkit:humanizeDetect and remove AI-generated writing patterns from text
/shipkit:onboardMulti-phase codebase onboarding
/shipkit:explain-systemExplore codebase and write verified system design docs
/shipkit:walkthroughTrace one feature end-to-end
/shipkit:update-rulesUpdate CLAUDE.md rules (never edit manually)
/shipkit:context-auditCheck context window health and find bloat
/shipkit:use-libraryRead docs before using any library
/shipkit:ai-featureScaffold AI/LLM features (chat, embeddings, RAG, agents)
/shipkit:legacy-auditAudit legacy codebase for modernization
/shipkit:migration-planPlan major dependency upgrades
/shipkit:ui-uxEmpathy-driven UI/UX design, review, and audit (web + mobile)

5 agents — subagents that do heavy work in their own context so yours stays thin:

AgentWhat It Does
grandfatherAnswers questions about one project (architecture, where-things-live, why) — reads its PROJECT_MAP.md, verifies against live source, returns a tight cited answer
eveAnswers questions across all your registered projects (the 360° view)
archivistBuilds/refreshes the PROJECT_MAP.md that grandfather and eve read
test-analyzerAuto-diagnoses test failures
codebase-explorerRead-only exploration: traces flows, maps architecture

The project elders — ask, don't pollute

The biggest idea in this release: instead of loading project knowledge into your main session (and burning context), you ask an elder and get back only the answer.

/shipkit:map                      # in a project: build its PROJECT_MAP.md once
/shipkit:map --register           # also add it to your cross-project registry
/shipkit:ask how does locale fallback work here?      # → grandfather (this project)
/shipkit:ask --all which apps deploy to Hetzner?      # → eve (all projects)

grandfather/eve do all the file reading in their own context and hand back a short, cited answer — your main context never sees the 40 file reads behind it. Use them for research, architecture questions, and "where/why" lookups. Keep editing in your main session. See GUIDE.md for details, and Episodic memory below for the optional decision-recall add-on.

2 knowledge bases — on-demand reference material:

KBWhat It Provides
code-review-standards8 review lenses, anti-pattern catalog, severity definitions
ui-ux-standardsCross-platform UI/UX: a11y, design, performance, mobile patterns

6 path-scoped rules — auto-loaded when editing matching files:

RuleTriggers On
testing.mdTest files (*_test.*, *_spec.*)
migrations.mdDatabase migrations
security.mdControllers, API, auth files
dependencies.mdDependency files (Gemfile, package.json, go.mod, etc.)
monorepo.mdMonorepo configs and workspace files
ui-ux.mdUI files (web, iOS, Android, Flutter, React Native)

Optional: Full Stack Configuration

Run /shipkit:setup to tailor the plugin to your project. It backs up your existing files first, and everything it does can be reversed with /shipkit:unsetup.

  1. Snapshots your current state — copies CLAUDE.md + .claude/ to .shipkit-backup-<timestamp>/
  2. Auto-detects your stack, test framework, and package manager
  3. Creates CLAUDE.md with your project info and workflow rules
  4. Installs stack-specific skills, rules, and knowledge bases
  5. Creates settings.json with safe permission defaults

Stack-Specific Additions

StackExtra SkillsExtra RulesExtra KBs
Rails/new-feature, /release, /safety-check, /deploy-checkgemfile, railscode-review-standards-rails, ai-rails
React/componentpackage-json, react—
Python/new-featurepyproject, python—
Go/new-featurego-mod, go—
Elixir/new-featuremix-deps, elixir—
Static/audit——

Optional: Episodic memory (MemPalace)

PROJECT_MAP.md captures structure — how a project is built today. It does not capture the narrative of decisions: "why did we pick Paddle over Stripe?", "what did we decide last session?". That history lives in your past conversations.

MemPalace is an optional, local-first memory store (MIT, no API calls) that indexes your conversation history verbatim and retrieves it with semantic search. ShipKit's grandfather and eve agents will automatically use it for decision-history questions if it is installed — and run perfectly fine without it (those questions just fall back to git history).

ShipKit does not bundle or auto-install MemPalace — it is a separate Python package plus a ~300 MB local embedding model, so it stays opt-in. To enable it:

# 1. Install MemPalace (puts `mempalace-mcp` on your PATH; uv or pipx)
uv tool install mempalace        # or: pipx install mempalace

# 2. Register it once at user scope so the elder agents can reach it
claude mcp add --scope user mempalace mempalace-mcp

# 3. Backfill a project's decision history from your Claude Code transcripts.
#    Transcripts are keyed by the directory you ran Claude in, under ~/.claude/projects/
mempalace mine ~/.claude/projects/-<your-project-dir> --mode convos --wing <project> --dry-run
mempalace mine ~/.claude/projects/-<your-project-dir> --mode convos --wing <project>   # for real

Restart Claude Code after step 2 so the server loads. The grandfather/eve agents already allowlist the mcp__mempalace__* tools, so only those two subagents can use them — and because Claude Code defers tool schemas by default (tool search), the ~30 MemPalace tools cost your main session almost nothing until an elder actually calls one. See GUIDE.md → Episodic memory for wings/rooms, repair, and the recall-is-a-claim caveat.

If you do not install MemPalace, nothing breaks — the elders simply skip the decision-recall step.

Uninstalling

Plugin-only users (never ran /setup): just uninstall the plugin. Nothing was written to your project.

/plugin uninstall shipkit@shipkit

Users who ran /setup: run /shipkit:unsetup first to restore your project, then uninstall the plugin.

/shipkit:unsetup          # restores CLAUDE.md and .claude/ from .shipkit-backup-<timestamp>/
/plugin uninstall shipkit@shipkit

/setup snapshots your entire CLAUDE.md and .claude/ directory to .shipkit-backup-<timestamp>/ before making any changes. /unsetup restores from that snapshot — your project goes back to exactly how it was.

Research

Design decisions in this plugin are informed by:

  • Do Context Files Actually Work? (ETH Zurich, 2025) — LLM-generated context files hurt performance. Human-written help only marginally. "Describe only minimal requirements."
  • Optimizing Coding Agent Rules (Arize, 2025) — Optimized rulesets contain 20–50 rules. Best rules are root-cause focused, correctness-preserving, edge-case aware.
  • Writing a Good CLAUDE.md (HumanLayer, 2025) — ~150–200 instruction limit for frontier LLMs. Progressive disclosure over monolithic files.

License

MIT


Support

If this project saves you time, consider sponsoring. It keeps development going and lets me know people are finding it useful.

  

Made with ❤️ by Anjan

Ver o original no GitHub

Bom para

Relacionados