firstintent

Getting started: Create a wallet and send your first transaction.

web3 balance: Check native token and ERC-20 token balances.

web3 send: Send native tokens or ERC-20 tokens to an address.

web3 CLI: Shared patterns for key management, global flags, and security rules.

web3 wallet: Create, import, show, export, and reset wallets.

本工作区级别的 Telegram 访问控制，覆盖全局 telegram:access skill。 操作本工作区的 .claude/channels/telegram/access.json，而非全局路径。 触发：/telegram:access pair <code>、pair <code>、approve、allowlist 等。

本工作区级 Telegram 接入，覆写全局 telegram:configure。 写入 .claude/channels/telegram/.env 与 .claude/settings.local.json 里的 TELEGRAM_STATE_DIR，绝不碰用户的 settings.json。 触发：接入 telegram、setup telegram、配 telegram bot、 /telegram:configure、token 是 <...>。

Send a Claude Code slash command to the current tmux pane. Use when someone says "/clear", "/compact", "/memory", "/new ...", "/review", "/help", "send /X", "execute /X here", "在当前 session 执行 /X", or pastes any slash command for the same Claude to run. ONLY works when the calling Claude is itself running inside tmux — refuses otherwise.

Render the current tmux pane to a PNG and reply with it via Telegram. Use when the user asks for a "screenshot", "snapshot", "看一眼现在的屏幕", "截图", "current pane", "current screen", or wants to see what Claude is showing in their terminal right now. ONLY works when the calling Claude is itself running inside tmux — refuses otherwise.

Delegate a long-tail investigation (log dig, audit, transcript analysis) to a peer agent and take back only a summary. Use when the raw output would crowd out 1000+ tokens of the primary session that do not need to live there.

Spawn N peer agents on genuinely independent subtasks, wait for every result, then merge. Use when subtasks do not read each other's outputs and the wall-clock saving is worth N× the token cost.

Delegate a verification check to a peer agent and consume a structured pass/fail/needs-info verdict instead of free-form text. Use when you want a second agent to check work against explicit criteria without rewriting it.

让两个或多个 vendor(如 Claude 和 Codex)并行做同一个任务,再盲评对比、推荐更优的一方。当用户想让不同引擎/模型 PK、要第二种实现、或对同一任务 A/B 两种方案时使用。

把一个 role / prompt / 工作流对着你定义的 benchmark 可量化地越改越好:变异 → 评测 → 多裁判打分 → 赢则保留、输则回滚,在护栏内(变异上限 / 停止条件 / 证据门槛)迭代,可无人值守过夜跑。当想让某个 agent / 角色在能打分的任务上自我改进(而非只重试)时用;前提是先有可机器打分的 eval。