--- name: autoclaude description: This skill should be used when the user asks to "run autoclaude", "automation", "one-shot", "single prompt", "batch automation", "loop automation", uses {autoclaude}, {oneshot}, {one-shot}, or needs Claude Code CLI automation (loop or one-shot). version: 1.0.0 user-invocable: true disable-model-invocation: false instructions_source: .claude\/skills\/autoclaude\/readme-instructions-autoclaude-claude-code-automation.md output_language: Vietnamese --- ## Autoclaude - Claude Code CLI Automation Hai mode thực thi: - **Loop mode**: Chạy N lần theo init file (LLP pattern). Cho batch tasks. - **One-shot mode**: Chạy 1 lần với prompt trực tiếp. Cho test nhanh. ## Mode 1: Loop (LLP - Linear Loop Prompt) Chạy Claude CLI N lần, mỗi lần đọc task tiếp theo từ init file. ### Workflow ``` init.json (tasks) → autoclaude-block.sh → build 3-layer prompt → claude CLI × N ``` ### Commands ```bash # Non-block (gnome-terminal, background) .claude\/skills\/autoclaude\/scripts\/autoclaude-nonblock.sh "\/path\/init.json" 30 [--cheap] # Block (inline, sequential) .claude\/skills\/autoclaude\/scripts\/autoclaude-block.sh "\/path\/init.json" 30 [--cheap] # List sessions .claude\/skills\/autoclaude\/scripts\/list-sessions.sh # Kill session .claude\/skills\/autoclaude\/scripts\/kill-stop-autoclaude.sh [SESSION_ID] ``` ### Flags (Loop mode) | Flag | Default | Description | |------|---------|-------------| | Position 1 | (required) | Init file path | | Position 2 | 1 | Loop count | | `--cheap` | opus | Use haiku model | | `--delay N` | 5 | Seconds between loops | | `--max-time N` | 600 | Timeout per loop (seconds) | | `--summary TEXT` | init basename | Task summary for Telegram | | `--test-retry` | false | Fast retry delays (1s,2s,3s) | ## Mode 2: One-shot (Direct Prompt) Chạy Claude CLI 1 lần duy nhất. Prompt trực tiếp từ CLI, không cần init file. ### Workflow ``` "prompt string" → autoclaude-oneshot.sh → wrap prefix\/suffix → claude CLI × 1 ``` ### Commands ```bash # Direct prompt .claude\/skills\/autoclaude\/scripts\/autoclaude-oneshot.sh "Check PHP for SQL injection" # With model .claude\/skills\/autoclaude\/scripts\/autoclaude-oneshot.sh "Fix login bug" --cheap # From file .claude\/skills\/autoclaude\/scripts\/autoclaude-oneshot.sh -f task.md # Raw (no prefix\/suffix wrapping) .claude\/skills\/autoclaude\/scripts\/autoclaude-oneshot.sh "Just test this" --raw # Silent (no Telegram) .claude\/skills\/autoclaude\/scripts\/autoclaude-oneshot.sh "Quick test" --no-telegram ``` ### Flags (One-shot mode) | Flag | Default | Description | |------|---------|-------------| | Position 1 | (required) | Prompt text | | `--cheap` | opus | Use haiku model | | `--raw` | false | No prefix\/suffix wrapping | | `--max-time N` | 600 | Timeout (seconds) | | `-f FILE` | - | Read prompt from file | | `--no-telegram` | false | Skip Telegram notifications | | `--save` | false | Save prompt to prompts\/ folder | ## Khi nào dùng mode nào | Tình huống | Mode | Lý do | |------------|------|-------| | Batch 10+ tasks | Loop | Init file quản lý tasks, auto mark done | | Test nhanh 1 prompt | One-shot | Không cần tạo file, chạy thẳng | | CI\/CD pipeline | One-shot + --raw | Prompt từ pipeline, không cần prefix | | Long-running project | Loop | Session tracking, retry, checkpoints | | Debug 1 step | One-shot + --cheap | Nhanh, rẻ | ## Architecture ``` .claude\/skills\/autoclaude\/ <- Skill home (scripts + docs) ├── SKILL.md <- This file ├── config.json <- SSoT config (flags, telegram, patterns) ├── scripts\/ <- All executable scripts │ ├── autoclaude-block.sh <- Loop mode (blocking) │ ├── autoclaude-nonblock.sh <- Loop mode (gnome-terminal) │ ├── autoclaude-oneshot.sh <- One-shot mode (NEW) │ ├── kill-stop-autoclaude.sh <- Kill sessions │ ├── list-sessions.sh <- List active sessions │ └── start-prompt.md <- Prompt template (loop mode) ├── docs\/ <- Documentation └── examples\/ <- Usage examples .fong\/claude-code-automation\/ <- Prompts storage ONLY └── prompts\/ <- Generated prompt files ``` ## Config (SSoT) File: `.claude\/skills\/autoclaude\/config.json` Key settings: - `claude_flags.dangerously_skip_permissions`: true (auto-approve) - `prompt_prefix`: [ultrathink] + mandatory rules - `prompt_suffix`: NO-QUIT-RULE + Telegram reporting - `telegram.token` + `telegram.chat_id`: Notification credentials - `patterns.linear-loop-prompt.checkpoint_interval`: 5 (git commit every 5 loops) ## Prompts Storage Prompts vẫn ở `.fong\/claude-code-automation\/prompts\/`. Không có symlinks. Chỉ prompts. ## References - `docs\/12-pattern-linear-loop-prompt.md` - LLP pattern chi tiet - `docs\/06-claude-cli.md` - Claude CLI flags - `docs\/07-config.md` - Config documentation - `docs\/08-telegram.md` - Telegram setup - `docs\/16-block-inline-mode.md` - Block vs non-block comparison - `readme-instructions-autoclaude-claude-code-automation.md` - Full instructions