create-general-agent

claude-agent-kit

A tiny kit for spinning up standalone personal AI agents with Claude Code (or any compatible CLI). It contains:

• create-general-agent — a skill that scaffolds one self-contained agent in its own folder, each with a two-part brain:
  1. Memory (memory/) — Claude Code-style file-based memory for durable facts, preferences, and feedback.
  2. Knowledge base (knowledge-base/) — Andrej Karpathy's "LLM wiki" method: an append-only raw/ folder (every source verbatim — the single source of truth) and a distilled wiki/ folder (encyclopedia-style articles with [[backlinks]] and an index) that the agent writes and continuously revises.
• tm-send — a hardened replacement for tmux send-keys for reliably delivering a message into another agent's tmux pane (optional; only if you run more than one agent).

Each created agent is standalone — it does not report to or coordinate with any other agent. It just works for you, in its own folder, and gets smarter over time by keeping its memory and wiki current.

Rule of thumb the agent follows: memory = what to remember about you and the work; wiki = what it has learned about the subject.

---

Install

Option A — as a Claude Code plugin (recommended)

In Claude Code:

/plugin marketplace add hungson175/claude-agent-kit
/plugin install create-general-agent@claude-agent-kit

Option B — copy the skill manually

git clone https://github.com/hungson175/claude-agent-kit.git
cp -r claude-agent-kit/skills/create-general-agent ~/.claude/skills/

Install tm-send (optional, only for multi-agent messaging)

cd claude-agent-kit
./bin/install-tm-send.sh          # installs to ~/.local/bin (override: ./bin/install-tm-send.sh /custom/bin)
tm-send --help

---

Use

Just tell Claude Code what you want the agent for — the skill triggers on phrases like "create an agent", "make me a tutor agent", "scaffold an agent with memory and a knowledge base". Or run the scaffolder directly:

~/.claude/skills/create-general-agent/scripts/create_general_agent.sh \
  rust-tutor "Teach me Rust by building a real CLI project."

cd ~/agents/rust-tutor && claude

Common options:

--project-dir PATH    where to create it (default ~/agents/<slug>)
--command CMD         CLI the agent runs in: claude (default), codex, opencode, ...
--context-file PATH   a file of extra mission context, embedded for first boot
--launch              also start it in a tmux session
--no-git              skip git init

What gets created

<agent>/
├── CLAUDE.md  +  AGENTS.md -> CLAUDE.md   # mission + brain rules + boot sequence
├── .gitignore
├── docs/  tasks/  outputs/  scripts/  context/
├── memory/                                # ① fast tier
│   ├── MEMORY.md                          #    index, loaded every session
│   └── _template.md                       #    one fact per file, with frontmatter
└── knowledge-base/                        # ② deep tier (Karpathy LLM-wiki)
    ├── README.md                          #    the update loop
    ├── raw/                               #    APPEND-ONLY sources (never edited)
    └── wiki/
        ├── index.md                       #    map of all articles
        └── _templates/article.md          #    distilled-article format

The knowledge-base loop (Karpathy's method)

Whenever the agent ingests something new:

1. Drop the source into knowledge-base/raw/ verbatim, with a short provenance header. Never edit raw files.
2. Read it, decide which wiki/ concept(s) it touches.
3. Create or update the relevant wiki/<concept>.md: revise the summary, fold in new facts, flag contradictions, strengthen [[backlinks]], and update wiki/index.md.

The raw/ pile only grows; the wiki/ gets smarter and more cross-linked. A memory file may also point into the wiki with [[a-wiki-page]].

tm-send quick reference

tm-send <role> "message"               # send to a role's pane in the current session
tm-send -s <session> <role> "message"  # target a specific session
tm-send --list                         # list roles (panes) in the session

It verifies the target pane belongs to the right session and handles per-CLI submit quirks (Claude Code / OpenCode need a second Enter; Codex needs a buffer-paste + delayed Enter) — the failure modes that make raw tmux send-keys drop or strand messages.

---

MIT licensed. Contributions welcome.