agentic-bootstrap

Vibe-coding architect for Claude Code. Investigates Anthropic's docs, the Claude Code ecosystem, and your codebase — then designs and builds the optimal agentic configuration for your project.

/agentic-bootstrap:init bootstraps. /agentic-bootstrap:refresh keeps it current.

What it does

• Discovers, doesn't hardcode. Investigates what mechanisms Claude Code currently supports (skills, MCPs, hooks, rules, subagents, workflows, or whatever exists today). Nothing about the platform is baked in.
• Tailors to your project. Detects your tech stack via Claude Code's built-in /init, then runs tech-specific investigation (Postgres MCPs, TypeScript skills, Payload-CMS plugins, etc.) plus pattern-finding queries for how to organize agentic config for your kind of project.
• Decides architecture first, tools second. Each recommendation carries what / how / why with sources. You see the architectural reasoning before any installation.
• Generates a rules tree. Instead of one monolithic CLAUDE.md, you get a .claude/rules/ directory with topic-organized files (code-quality, testing, security, plus tech-specific files per detected stack).
• (Optional) deep mode. Opt into LLM-driven code analysis — the skill reads your code, asks clarifying questions when uncertain, then drafts custom skills/hooks/rules for gaps no existing tool fills.
• Refresh-safe. Tracks file hashes; respects your manual edits; pauses for keep-yours / take-upstream / merge-by-hand decisions.
• MCPs gated by explicit consent. Each MCP install confirms capabilities before it runs. No silent installs.

Install

Requirements

• Claude Code v2.1.110+ (for plugin dependency resolution — this plugin declares dependencies on claude-md-management and other official plugins)

Plugin marketplace (when this plugin is published)

/plugin install agentic-bootstrap

Local install for development or testing

Clone the repo somewhere, then point Claude Code at it with --plugin-dir:

git clone https://github.com/petarjakopec/claude-agentic-bootstrap.git
claude --plugin-dir ./agentic-bootstrap

While a session is running, use /reload-plugins to pick up any edits without restarting.

Plugin skills are namespaced, so the commands become /agentic-bootstrap:init and /agentic-bootstrap:refresh.

Use

In any project directory:

/agentic-bootstrap:init      # first-time bootstrap
/agentic-bootstrap:refresh   # update an existing setup

Both flow interactively. Sensible defaults; press Enter to accept. Explicit confirmation required for MCP installs.

What gets created in your project

.claude/
├── CLAUDE.md                    ← thin entry; imports rules
├── agentic/
│   ├── sources.yaml             ← you-editable: extras, force-install, blacklist
│   ├── catalog.yaml             ← discovered items + provenance
│   └── state.yaml               ← installed items + file hashes
├── rules/                       ← smart defaults from project context
│   ├── code-quality.md
│   ├── testing.md
│   └── [tech-specific + project-type files]
├── skills/   plugins/   hooks/  ← installed
└── settings.json                ← MCP/hook configs

Project-only by design. The skill itself lives at ~/.claude/; everything it installs goes into the project. Commit .claude/ to share setup with teammates, or .gitignore it for personal config.

Customizing investigation

Edit .claude/agentic/sources.yaml directly, or use prompts during /agentic-bootstrap:refresh:

sources_extra:
  - url: https://github.com/some/cool-claude-skills
    note: "Recommended by colleague"
forced_install:
  - name: postgres-query-reviewer
    type: skill
    reason: "Team standard"
blacklist:
  - name: noisy-skill
    reason: "Too many false positives"
pinned:
  - name: code-quality-baseline
    type: skill

How it works

See docs/DESIGN.md for the full design (investigation model, source layers, init/refresh flows, hallucination guards, edit detection).

Status

v0.1.0 — early prototype. Designed and drafted; not yet battle-tested. Expect rough edges around /init integration, plugin marketplace mechanics, and deep-mode prompts. Feedback via issues welcome.

License

MIT. See LICENSE.