delilah-code-setup
My personal Claude Code setup. Minimal, intentional, no bloat.
Built around four tools that work together: semantic doc search, token-efficient shell commands, code graph navigation, and spec-driven feature development.
Tools
| Tool | What it does | Why |
|---|
| QMD | On-device semantic search over your docs | Ask Claude about architecture decisions, past sessions, or any markdown — searches locally, no API calls |
| RTK | Rewrites shell commands transparently to cut token usage 60–90% | git status output drops from thousands of tokens to hundreds. Happens via hook, zero friction |
| Codegraph | AST-based code knowledge graph exposed as MCP tools | Claude navigates code via graph queries instead of reading files. 70% fewer tool calls, 59% fewer tokens |
| Spec-kit | Spec-driven feature development | Define what before how. Generates spec → plan → tasks before touching code |
Install
Prerequisites
- macOS or Linux
- Claude Code CLI
- Bun — installed automatically if missing
- Homebrew (macOS only, for RTK) — fallback curl installer used on Linux
Option A — Claude Code plugin (skills + agents only)
In Claude Code, run the slash command to register the marketplace:
/plugin marketplace add FranPerezFolgado/delilah-code-setup
Then install:
claude plugin install delilah-code-setup@FranPerezFolgado
After that, run the setup script to install the system tools (QMD, RTK, Codegraph, spec-kit):
curl -fsSL https://raw.githubusercontent.com/FranPerezFolgado/delilah-code-setup/main/setup.sh | zsh
Option B — Full install script (tools + skills + agents)
git clone https://github.com/yourusername/delilah-code-setup
cd delilah-code-setup
chmod +x setup.sh && ./setup.sh
Installs QMD, RTK, Codegraph, spec-kit, configures Claude Code globally, starts the QMD daemon, and copies all skills and agents. Restart Claude Code when it finishes.
What gets installed
~/.claude/
├── CLAUDE.md ← global instructions (QMD + Codegraph + RTK)
├── settings.json ← MCP servers, RTK hook, auto-allow permissions
├── skills/
│ ├── recall.md ← /recall
│ ├── feature.md ← /feature
│ ├── review.md ← /review
│ ├── pr-review.md ← /pr-review
│ ├── commit.md ← /commit
│ ├── doc-check.md ← /doc-check
│ ├── code-nav.md ← /code-nav
│ ├── project-setup.md ← /project-setup
│ ├── setup-qmd.md ← /setup-qmd
│ ├── save-session.md ← /save-session
│ ├── adr-check.md ← /adr-check
│ └── stats.md ← /stats
└── agents/
├── code-researcher.md ← Haiku: code search via Codegraph
├── doc-searcher.md ← Haiku: QMD queries
├── commit-writer.md ← Haiku: commit message generation
├── doc-checker.md ← Haiku: stale doc detection
├── security-reviewer.md ← Sonnet: security review
├── code-quality-reviewer.md ← Sonnet: SOLID, patterns, correctness
├── test-reviewer.md ← Haiku: test coverage and quality
└── architecture-reviewer.md ← Sonnet: coupling and structure
# QMD daemon (auto-starts at login)
macOS → ~/Library/LaunchAgents/com.qmd.mcp.plist (launchd)
Linux → ~/.config/systemd/user/qmd-mcp.service (systemd)
Other → entry added to ~/.zshrc / ~/.bashrc
Agents
Subagents invoked by skills. Lightweight tasks use Haiku; review tasks use Sonnet.
| Agent | Model | Called by | What it does |
|---|
code-researcher | Haiku | /feature, /code-nav | Searches codebase via Codegraph — symbols, call chains, affected files |
doc-searcher | Haiku | /recall, /feature | Queries QMD collections with lex + vec search |
commit-writer | Haiku | /commit | Generates Conventional Commits message from staged diff |
doc-checker | Haiku | /doc-check | Finds documentation that's become stale after code changes |
security-reviewer | Sonnet | /review, /pr-review | OWASP, secrets, auth/authz, injections, data exposure |
code-quality-reviewer | Sonnet | /review, /pr-review | Correctness, SOLID, design patterns, error handling |
test-reviewer | Haiku | /review, /pr-review | Coverage, test quality, edge cases |
architecture-reviewer | Sonnet | /review, /pr-review | Coupling, consistency, breaking changes, abstractions |
Skills
Session
