A batteries-included working environment for Claude Code. Dispatch agents, skills, and statusline — all extracted from daily production use.

What's Inside

ai-toolkit/
├── agents/                  # SOP agents (symlink to ~/.claude/agents/)
│   ├── analyst.md           # Architecture / Planning / Audit
│   ├── investigator.md      # Search / Explore / Debug / External research
│   ├── builder.md           # Code implementation / Testing
│   ├── reviewer.md          # Code review / Dead code cleanup
│   ├── doc-sync.md          # Doc init / Doc sync
│   └── agent-factory.md     # Design and generate new agents
├── skills/                  # Skills (symlink to ~/.claude/skills/)
│   ├── handoff/             # Session handoff
│   ├── thorough/            # Relentless delivery mode
│   ├── vector-memory/       # Persistent vector memory usage guide
│   ├── project-docs/        # Project documentation structure
│   ├── agentskill-expertise/ # Skill design knowledge base
│   ├── collaboration-style/ # AI-human collaboration framework
│   └── self-growth/         # Continuous learning framework
├── mcp.example.json         # MCP server config template
└── statusline/              # Cost & context monitoring

Agents

SOP-style prompts for Claude Code's Agent tool. When /thorough dispatches parallel subagents, prompt quality determines output quality — these agents provide step-based workflows with hard thresholds, classification heuristics, and structured output formats.

Agent	Model	When to use
`analyst`	opus	"design this", "plan the implementation", "audit codebase health"
`investigator`	sonnet	"find all usages of X", "how does this work", "why does this fail"
`builder`	sonnet	"implement this", "modify the handler", "write tests for X"
`reviewer`	sonnet	"review this code", "find dead code", "clean up unused exports"
`doc-sync`	sonnet	"set up project docs", "sync docs after changes"
`agent-factory`	opus	"create a new agent", "improve this agent's prompt"

Each agent auto-detects its mode from dispatch context. One agent, multiple workflows.

Design Principles

Zero concept explanation — All operational instructions. Claude already knows what CQRS is.
Step-based SOP — "Do X, then Y, if Z threshold → action." Not "You are an expert at..."
Hard rules as threshold + trigger — >50 lines → flag, >4 nesting levels → flag. Not "keep functions small."
Classification heuristics — AUTO-FIX / ASK / CRITICAL with concrete criteria. Not checklists.
Structured output — Every agent ends with a report template. Consistent, parseable.

Skills

Skill	Description
`/handoff`	Session handoff — compress context into a structured prompt for seamless continuation
`/thorough`	Relentless delivery mode — exhaust all options, cost-aware model selection, verify before done
`/vector-memory`	Persistent vector memory via LanceDB — store facts, decisions, lessons across sessions
`/project-docs`	Project documentation structure — standard `proj-[name]/` layout with ADRs, stories, and operations guides
`/agentskill-expertise`	Agent Skill design knowledge base — mechanisms, philosophy, patterns, pitfalls
`/collaboration-style`	AI-human collaboration norms — friction cases, coding style, behavioral guidelines
`/self-growth`	Continuous learning framework — learn from work, organize knowledge, build feedback loops

MCP Servers

Example configuration for the MCP servers used in this toolkit.

mcp.example.json — copy to your project as .mcp.json and fill in your API keys.

Server	What it does
`@cablate/memory-lancedb-mcp`	Persistent vector memory with hybrid search (semantic + keyword)
Serena	Semantic code intelligence — symbol search, references, refactoring

Statusline

Cost and context monitoring for Claude Code. Two-line display with context alerts and plan usage tracking.

 Normal (< 60% context):
┌──────────────────────────────────────────────────────────────────┐
│ Claude Opus 4  | [=======--------------] 45.2K/200.0K 22.6%    │
│ 5h: 12.3% (4h 22m) | 7d: 8.1% (6d 3h)                        │
└──────────────────────────────────────────────────────────────────┘

 Warning (>= 60% context):
┌──────────────────────────────────────────────────────────────────┐
│ Claude Sonnet 4 | concise | [============--------] 130.5K/200.0K 65.3%  /handoff soon │
│ 5h: 45.0% (2h 10m) | 7d: 22.4% (5d 1h)                       │
└──────────────────────────────────────────────────────────────────┘

 Critical (>= 80% context):
┌──────────────────────────────────────────────────────────────────┐
│ Claude Opus 4  | [==================--] 310.0K/200.0K 95.0%  !! HANDOFF NOW !! │
│ 5h: 78.2% (1h 05m) | 7d: 51.3% (3d 12h)                      │
│ !! DO NOT close/resume -- use /handoff first, or waste 6%+ of 5h tokens !! │
└──────────────────────────────────────────────────────────────────┘

Line 1 — Model name, output style (if not default), context progress bar with K-precision token counts, usage %, and alerts at 150K/200K/300K thresholds.

Line 2 — 5-hour and 7-day plan usage rates with reset countdowns. Fetched from Claude API (cached 5min) or inline rate_limits (v2.1.80+).

Line 3 — Appears at 250K+ tokens. Hard warning against closing/resuming without handoff.

statusline/statusline.ps1

// ~/.claude/settings.json
{ "status_line_command": "powershell -NoProfile -File C:/Users/YOU/.claude/statusline.ps1" }

License

MIT