360 Secure SkillHub
360 安全技能中心

Task Extractor

核心功能:从用户含3+行动项的消息提取任务,生成结构化队列,按依赖+优先级执行,实时更新状态并反馈进度,防止任务遗漏。典型场景:用户发送多指令混合请求、批量任务消息。输入为用户多任务文本,输出为任务清单、进度报告、最终完成情况,同步生成TASK_QUEUE.md作为任务管理源。

dodge1218
dodge1218
热度 33.00
最近更新 2026-05-11 17:38:33
分类:效率依赖:读写文件 / 内部调用大模型应用场景:任务管理 / 工作流

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add task-extractor
Metadata
name
task-extractor
version
1.0.0
description
Extract, track, and verify completion of multiple tasks from a single user message. Use when any message contains 3+ actionable items, a prompt dump with mixed instructions, or a compound request. Prevents task drop by saving to TASK_QUEUE.md before executing. Reports completion status per item.

Task Extractor

Parse multi-task messages → numbered queue → execute sequentially → verify each → report.

Activation

Trigger on ANY user message containing 3+ distinct actionable items. Signs:

  • Multiple sentences starting with verbs ("build", "fix", "send", "check", "add", "research")
  • Comma-separated or period-separated instructions
  • Mixed topics in one paragraph
  • "also", "and then", "plus", "oh and", "one more thing"

If uncertain whether to activate: activate. False positives (structuring a 2-item request) cost nothing. False negatives (dropping task #7 of 12) cost trust.

Step 1: EXTRACT (before any execution)

Parse the message into individual tasks. Each task gets:

  • Number (sequential)
  • Summary (one line, imperative verb)
  • Type: BUILD | FIX | RESEARCH | SEND | DEPLOY | CONFIG | OTHER
  • Estimated effort: QUICK (< 5 min) | MEDIUM (5-30 min) | HEAVY (30+ min / sub-agent)

Write to workspace/TASK_QUEUE.md:

# TASK_QUEUE — [date] [time]
# Source: [channel]
# Total: [N]
# Status: IN PROGRESS

| # | Task | Type | Effort | Status | Artifact |
|---|------|------|--------|--------|----------|
| 1 | [summary] | BUILD | HEAVY | ⏳ | |
| 2 | [summary] | FIX | QUICK | ⏳ | |
| 3 | [summary] | SEND | QUICK | ⏳ | |
...

Step 2: RECEIPT

Reply to the user with the extracted checklist BEFORE starting work:

📋 Extracted [N] tasks from your message:

1. ⏳ [task summary]
2. ⏳ [task summary]
3. ⏳ [task summary]
...

Starting now. I'll check each off as I go.

Do NOT ask "is this right?" unless genuinely ambiguous. Convert ambiguity into a task and execute. Ryan's rule: don't interrogate, just do.

Step 3: EXECUTE

Work through tasks in dependency order (not necessarily numerical order):

  • Independent QUICK tasks first (batch them in parallel tool calls)
  • MEDIUM tasks next
  • HEAVY tasks: spawn sub-agents with clear scope

For each completed task, update TASK_QUEUE.md:

  • ⏳ → ✅ (done) or ❌ (failed) or ⚠️ (partial) or 🔄 (spawned sub-agent)
  • Fill in the Artifact column (file path, URL, commit hash, or "n/a")

Step 4: RECONCILE

After all tasks are attempted (or sub-agents spawned), reply with the final checklist:

📋 Task Report ([completed]/[total]):

1. ✅ [task] → [artifact]
2. ✅ [task] → [artifact]
3. 🔄 [task] → sub-agent running, will announce when done
4. ❌ [task] → [reason for failure]
5. ⚠️ [task] → [what was done, what's left]

Step 5: VERIFY (on sub-agent completion)

When a sub-agent announces completion:

  • Update TASK_QUEUE.md
  • If ALL tasks are now ✅/❌, send final summary
  • If tasks remain ⏳, continue working

Rules

  1. NEVER skip the extract step. Even if the answer seems obvious. The extract step IS the safety net.
  2. NEVER mark a task ✅ without evidence. Evidence = file exists, command succeeded, API returned 200, deploy URL works.
  3. If a task fails, say WHY. Not just ❌ — include the error, the blocker, or what's needed.
  4. Sub-agent tasks get 🔄 until completion event arrives. Don't mark them ✅ optimistically.
  5. TASK_QUEUE.md is the source of truth. If context overflows, re-read it. If a session restarts, re-read it.
  6. One TASK_QUEUE at a time. If a new multi-task message arrives while one is active, append to the existing queue (renumber).
  7. Checkpoint at 5 completed tasks. Update the user with progress so far.

Edge Cases

"Do X and also Y but wait on Z"

  • X and Y get ⏳, Z gets 🕐 (blocked — note the dependency)

Sub-agent timeout

  • Mark as ⚠️, note what was attempted, offer to retry or take over manually

User changes mind mid-execution

  • Update TASK_QUEUE.md, cross out cancelled tasks with strikethrough, continue with remaining

Overlapping tasks

  • If task 3 and task 7 are really the same thing, merge them. Note in Artifact: "merged with #3"

Anti-patterns

  • ❌ Reading the message, immediately jumping into task #1 without extracting all tasks
  • ❌ Marking a sub-agent task ✅ before the completion event
  • ❌ Saying "I'll get to that" and then forgetting
  • ❌ Only reporting on the tasks you completed, silently dropping the ones you didn't
  • ❌ Asking "which one should I do first?" — just prioritize by dependency + effort and go