AIgent Operator Kit

Five specialists. One install. Drop them into ~/.claude/agents/.

Five Claude Code subagents that each do one job well (design, build, scout, research, critique), plus a context-loader hook that stops you re-explaining your project every session. MIT licensed. Use them as-is or adapt them.

Install

One line. Backs up your settings.json, installs the agents, and wires the context-loader hook (idempotent, never clobbers existing hooks):

curl -fsSL https://raw.githubusercontent.com/wrg32786/operator-kit/main/install.sh | bash

Prefer to do it by hand? Just copy the agents:

cp agents/*.md ~/.claude/agents/

Claude Code auto-discovers agents in that directory. Restart it, then invoke one:

use the echo agent to find every place we call the Stripe API

The agents

One agent per job beats one generalist at average quality. Compose them: Echo finds the files, Newton researches the approach, Hypatia challenges it, Lyra builds it.

Two patterns worth calling out. Lyra and Newton end every response with an honesty ledger: what changed, what was left alone, what was noticed but not fixed, what is still uncertain. You stay informed even when you are moving fast. Iris writes specs, not code: a brief precise enough that a builder implements without guessing.

The context loader

A UserPromptSubmit hook. Mention a keyword you have configured ("auth flow", "payments") and the relevant project files are injected into context before the agent responds. Structural enforcement, not a memory note: it fires every time, not just when you remember to.

// project-keywords.json
"auth": {
  "keywords": ["auth flow", "login", "session"],
  "priority_file": "docs/auth.md",
  "files": ["docs/auth.md", "src/lib/session.ts"]
}

The installer drops a starter keywords file at ~/.claude/hooks/operator-kit-keywords.json. Point its entries at your files, restart, and type a keyword to see the [AUTO-CONTEXT] block appear. Full walkthrough in context-loader/install.md.

The critical-rules template

rules/post-compact-critical.md.template is a starting point for rules that must survive a long session. When Claude Code compacts, most history is lost; files wired into your project settings do not. Capture the things that cannot drift: database invariants, style rules, known footguns, verification gates.

What's in it

operator-kit/
├── agents/          # iris · lyra · echo · newton · hypatia  → ~/.claude/agents/
├── context-loader/  # auto-context-load.sh + keywords + install guide
├── rules/           # post-compact rules template
├── examples/        # a filled-out keywords file
└── install.sh       # one-click installer

License

MIT. See LICENSE. The agents are self-contained: no private skills, no hidden dependencies, nothing to phone home.