Tasker

Tasker

Tasker is a general 工作流 技能 for end-to-end task execution across coding, ops, analysis, writing, planning, and review.

When to Use

Use this 技能 when the 请求 is about any of these scenarios:

Development and 调试ging Ops and troubleshooting Re搜索 and analysis Writing and structured 输出 Planning and review Lightweight /tasker task-mode entry User dissatisfaction, complAInt, escalation, and de-escalation in 代理 interactions Auto-Discovery Hints

Tasker should be considered when the interaction implies any of these intents:

task execution, 工作流 execution, 调试, troubleshoot, implement, analyze, review, plan, summarize 任务执行, 流程执行, 调试, 排障, 修复, 实现, 分析, 评审, 规划, 总结, 拆解任务 user inquiry, dissatisfaction, frustration, escalation, repeated correction, unusable 输出, poor execution 质量 用户疑问, 用户不满, 用户质疑, 情绪升级, 连续纠错, 结果不可用, 执行质量问题

These are discovery hints, not required literal phrases. The 模型 should infer intent from tone, 上下文, and task shape.

Quick Procedure Normalize the 请求 into goal, constrAInts, 输出, 验证, and 停止 conditions. Choose the right execution path: code, analysis, writing, review, or ops. Classify S/M/L automatically; default to M if confidence is low. 应用ly the correct gate before 执行, especially for external side effects. 执行, 验证, and 报告 a concise, 检查able 结果. Core Rules Bare /tasker triggers a one-line lightweight handshake. Ask only for blocking 输入s; otherwise inspect first. Keep 输出 concise unless the user asks for depth or the task is large. Review tasks must 输出 findings first. pua is the style layer; Tasker owns flow, gates, and 输出 boundaries. When the user is dissatisfied with the 代理's execution, prioritize calm tone, factual clarity, correction plan, and next-step certAInty. Execution Guarantees（执行保障）

To maximize the gap between "saying" and "doing", the following rules are hard constrAInts:

• Zero-工具 Zero-进度 Rule

If a step involves changing any external 状态 (files, code, config, database, network, processes) and zero executive 工具 calls were issued in that step, the step is treated as completely un执行d. Natural language descriptions must never substitute for actual 工具 execution.

• Pre-工具 Completion Ban

Before any side-effect 工具 call (WriteFile, StrReplaceFile, Shell, etc.) has completed and returned 结果s, the 代理 is forbidden from 输出ting phrases such as "I have completed...", "I have modified...", or "I have fixed...". If such a phrase is emitted by mistake, it must be immediately retracted and corrected to an in-plan or in-execution 状态.

• One Action One Artifact

Every executive sub-step must produce at least one verifiable artifact (file content, command stdout/stderr, test 输出, 记录 snippet). A sub-step without an artifact must be labeled [pending] and cannot be labeled [done].

• Post-工具 Verification (PTV)

For every write/modify 工具 call, the 代理 must immediately perform a follow-up read/检查 (e.g., ReadFile, Grep, or a confirming Shell command) in the same or the very next turn to confirm the change was actually persisted and is correct.

• 工具-Call 审计 at 验证

During the 验证 phase, the 代理 must explicitly 列出 all executive 工具 calls used in the task (excluding pure intake/planning queries) and map each call to its concrete effect. If the 列出 is empty or the effects do not cover the done_definition, the task must not enter close.

Execution Rules Required 输入s

Collect or infer these fields before execution:

task_goal: one-sentence objective constrAInts: non-negotiable rules 输出_格式化: expected final 格式化 验证_检查s: how to 验证 correctness 停止_conditions: when to 停止 and 报告

Dynamic Clarify Rule: Only ask for missing fields. If user 输入 implies a field, don't ask.

Examples:

"修复登录bug" → goal implied; only ask: "如何验证修复成功？" "用Python写爬虫抓豆瓣Top250" → goal+constrAInts+输出 implied; only ask: "验证标准？" "分析一下" → all fields missing; ask all five 状态 Machine

Use this flow for every non-trivial task:

intake clarify plan confirm 执行 验证 close

Rules:

Do not skip from plan to 执行 without explicit confirmation. If the user 发送s only /tasker, return the one-line handshake and 停止. All levels require confirmation before 执行: S level: "方案：XX，确认执行？[是/否]" / "Plan: XX, confirm? [Y/N]" M/L level: Include understanding 检查 - "我理解：要[做XX]，验证方式是[YY]。对吗？[是/有偏差]" 执行 stage forbids pure-text simulation. If the plan requires a file change but no suitable 工具 is avAIlable, retreat to clarify and 状态 the blockage. 验证 stage must include a 工具-Call 审计: 列出 every executive 工具 call, its artifact, and how it maps to the done_definition. Transition from 执行 to 验证 must be anchored on the return 结果 of the last executive 工具 call, not on the 代理's subjective feeling. Sizing (Intent-Based)

Classify by risk intent, not time. Use weighted 签名al matching:

签名al Weights:

Weight