Skip to content

Tools cheat sheet

Five tools, one pattern. Paths and features change monthly; check /help in the version you install.

Claude Code OpenCode Codex CLI Cursor Kilo Code
Surface terminal, VS Code, desktop, web terminal TUI, VS Code terminal, VS Code, web IDE (VS Code fork) + CLI VS Code / JetBrains extension + CLI
Models Anthropic any provider, incl. local OpenAI several, routed hundreds, bring your own key / local
Instructions CLAUDE.md (+ .claude/rules/) AGENTS.md (reads CLAUDE.md too) AGENTS.md (+ ~/.codex/) .cursor/rules/*.mdc rules + Memory Bank
Skills .claude/skills/*/SKILL.md .opencode/skills, .claude/skills, .agents/skills .codex/skills, .agents/skills own format custom modes
Memory auto-memory MEMORY.md + /memory AGENTS.md + session AGENTS.md + session codebase index Memory Bank (Markdown in repo)
Enforcement hooks + permission rules allow / ask / deny config sandbox + approval modes rules only rules only
Licence proprietary open source open source (CLI) proprietary open source (MIT)

Same move, three keystrokes

Move Claude Code Codex CLI OpenCode
Draft the instructions file /init → CLAUDE.md /init → AGENTS.md /init → AGENTS.md
Read-only planning Shift+Tab → plan mode /plan Tab → Plan agent
Undo the agent's changes Esc Esc → /rewind (code, chat or both) git; inspect with /diff /undo, /redo (git-backed)
Free up context /compact <what to keep>, /clear /compact, /new /compact, /new
Shell output into context !uv run pytest -x !uv run pytest -x !uv run pytest -x
Attach a file @path @path or /mention path @path
Reusable procedure .claude/skills/x/SKILL.md → /x .agents/skills/x/SKILL.md → $x .opencode/commands/x.md → /x
Pick up yesterday's session claude -c, /resume codex resume, /resume, /fork opencode -c, /sessions
Parallel sessions, one per worktree claude -w <name> -n <name> git worktree add, then codex in each git worktree add, then opencode in each; resume with opencode -s <ses_id>
Headless / CI claude -p "…" codex exec "…" opencode run "…"
No prompts at all (sandbox only) --dangerously-skip-permissions --yolo (or -a never -s workspace-write) --auto (deny rules still apply)

No-prompt modes

Run without approval prompts only inside a sandbox, container or scratch clone — never on a shared allocation or production data.