claude-skills — Repository Architecture

System overview

flowchart TD
  MP["marketplace.json
.claude-plugin/"]
  IOS["ios-dev
34 skills · 6 cmds"]
  LIN["linear-pm
1 skill · 6 cmds"]
  CW["core-workflow
6 skills · 4 cmds · 2 agents · 1 hook"]
  SW["second-wind
skill wrapper"]
  PC["prompt-craft
6 skills · 6 hooks"]
  WIND[("tools/second-wind
wind.py")]
  AD["adapters/install.sh"]
  SEED["setup/
setup.sh · capture.sh · contribute.sh"]
  CC(("Claude Code"))
  OTHER(("Codex / Copilot /
AGENTS.md agents"))

  MP --> IOS & LIN & CW & SW & PC
  SW -.->|"wraps"| WIND
  CC -->|"/plugin marketplace add"| MP
  SEED -->|"local_plugins step"| CC
  SEED -->|"wind → ~/.local/bin"| WIND
  AD -->|"symlink skills"| OTHER
  IOS & LIN & CW & SW & PC -->|"skills/"| AD

  classDef plugin fill:#0e8fb01a,stroke:#0e8fb0,stroke-width:2px
  classDef infra fill:#b07a181a,stroke:#b07a18,stroke-width:2px
  class IOS,LIN,CW,SW,PC plugin
  class SEED,AD infra

The five plugins

`ios-dev`

build → screenshot → phone → ship

app-preview — build on simulator, screenshot, deliver to iPhone (iMessage ping + iCloud Drive)
ios-build — sim/device builds with signing auto-detection
release — archive → .ipa → validate → upload to App Store Connect
/preview, /fix commands — screenshot to your iPhone in one shot

`linear-pm`

issue conventions + autonomous pickup

linear-pm skill — label vocabulary, status taxonomy, branch/PR naming, .claude/linear.yml policy
6 commands — init, new, pick (the autonomous loop), status, sync, block

`core-workflow`

everyday glue

commit — shellcheck-gated conventional commits
contribute — push changes back to this repo from anywhere
agents — image-parser (vision), web-researcher (docs lookup)
hook — shellcheck on every .sh Edit/Write

`second-wind`

outlast the 5-hour limit

wind CLI — one tmux session per repo, watches for the usage-limit pause, resumes everything after reset
skill wrapper — teaches the agent wind init/add/up/watch/status/resume/down + self-install fallback

`prompt-craft`

rough ask → deterministic spec

/improve-prompt — restated goal, acceptance criteria, assumptions, recommended commands
/plan — decompose into per-step acceptance criteria + a TodoWrite plan
debug / refactor / review lenses — correctness + security first
/refresh — re-scan skills and rebuild the command-advisor registry + history profile on demand; also wires the statusline
command advisor — ranks the most relevant next command by combining keyword relevance, your command-history frequency, and git-state context (at most 3 recommendations). It powers the pre-prompt banner, the Stop-hook follow-ups, and the statusline segment. Every surface is user-only — advisor output never enters the model context.
hooks — suggest_next (Stop, now routed through the command advisor), prompt_hint (UserPromptSubmit), registry_freshness (SessionStart), statusline_hint (statusLine) + opt-in block_secrets / format_on_edit

How installs flow

Public user — Claude Code

anyone, two commands

/plugin marketplace add abhijitbansal/claude-skills → /plugin install <name>@claude-skills → skills + commands + hooks live

/plugin marketplace add abhijitbansal/claude-skills

/plugin install core-workflow@claude-skills

Public user — other agents

clone, then one script

git clone → adapters/install.sh codex | copilot | agents-md → same SKILL.md files, symlinked or indexed

Owner — fresh machine seed

setup.sh, eight idempotent steps

preflight→ claude→ marketplaces→ plugins→ skills→ dotfiles→ local_plugins→ symlinks

One source of truthThe owner's machines install the repo's own plugins through the same marketplace mechanism the public uses (local_plugins step) — there is no second code path to drift.

Markdown twin: docs/architecture.md · Catalog: skills & tools catalog · Repo: github.com/abhijitbansal/claude-skills