# Agent Toolkit Guide A collection of reusable skills and rules for AI agents (Claude Code and similar). Everything must be self-contained and generic — reusable in any project — so avoid project-specific logic or assumptions. Rules are opt-in: a skill must work correctly with no rules installed, so never fix or extend a skill by adding a rule it depends on. `docs/core-philosophy.md` holds the project's five pillars and their litmus test; read it before creating or redesigning a skill, or when brainstorming toolkit direction. Frame rules and skills as agent- and model-agnostic: describe what they do, not tied to a specific agent (e.g. "for Claude Code") or model. Install instructions may still name agent-specific paths (e.g. `~/.claude/skills/`) — that's the install mechanism, not the content's framing. New skills ship at version `0.1` as a trial; bump to `1.0` only after a successful real-world run. In Markdown, wrap prose lines at the `max_line_length` in `.editorconfig`. Never break code (fenced blocks or inline backtick spans — a command stays on one line even past the limit), tables, URLs, links, or YAML frontmatter values to satisfy it. Skill test harnesses live in `test//` — repo-internal, never installed; their `runs/` output dirs are gitignored. When changing skills, rules, manifests, install behavior, or repository conventions, update the docs in the same change — `README.md`, `AGENTS.md`, `docs/`, and any affected artifact documentation — so a fresh agent session understands the current behavior without prior conversation context. Keep `README.md` lean: catalogs and quick-install commands only; detailed guides belong in `docs/`. ## Multi-agent quirks Skills with `disable-model-invocation: true` also need `type: flow`, or Kimi Code hides them even from manual invocation. Never put a `:` in a skill description — `: ` in the unquoted value breaks strict YAML parsers (e.g. Copilot's); avoid the character entirely instead of quoting.