auto-task · Autonomous Execution of Complex Long-Running Tasks

1. Trigger Judgment

✅ Suitable for auto-task

• The user has a complex or vague task that needs decomposition before execution
• The answer/output is not in the user's mind; the AI needs to run tools/research/analysis on its own to derive it
• Long time span (multi-step, multi-tool, may require sub-agent)
• The user explicitly or implicitly expects the AI to be self-driven

❌ Not Suitable (Route to Other Skills)

2. Core Principles (Hard Constraints)

1. The task queue is the single source of truth —— The §3 Task Queue in TASK.md is the state of the entire task; you must actively verify it after each batch execution
2. Triggered Reporting —— Do not disturb the user by default; only stop when any condition in §6 is met
3. Full Freedom of Tools —— Any tool can be used: WebSearch / WebFetch / Bash / Read / Write / Edit / Glob / Grep / Sub-agent / Any MCP / Parallel Calls / Background Tasks
4. Honest Declaration —— For testing/evaluation tasks, you must mark the following at the top of the output report: data source, whether it's a blind test, potential contamination
5. Plain Answers for Plain Questions —— Answer in plain language when the user asks colloquially; don't overuse structures and subheadings
6. Position Follows Evidence —— Update your position when new data comes in; don't just "supplement" your original position

3. Initiation Protocol (First-Round Setup)

Step 1 · Task Confirmation ⚠ Critical

• The core of the task I understand
• The boundaries I understand: what to do, what not to do
• The output format (report / decision / tool / document / list / ...)
• If there are unavoidable questions, at most 1-2 most critical ones

• ❌ Ask 5 detailed questions right away
• ❌ Break down the task queue right away
• ❌ Start executing right away

Step 2 · Decompose Task Queue

• ID: T1, T2, T3...
• Task Content: One-sentence description
• Expected Output: What should be obtained after completing this task
• Status: ⬜ Pending / 🔄 In Progress / ✅ Completed / ❌ Abandoned / ⏸ Waiting for Dependencies
• Dependencies: dep: T1, T3 (if any)
• Tool Suggestions: Optional, decided by the AI

• Too coarse → Loss of control, deviation from the task
• Too fine → Rigidity, no room to handle unexpected situations
• Experience rule: Each task should take 5-15 minutes to execute

Step 3 · Autonomy Mode (Default: b)

• (a) Conservative —— Report after completing N items
• (b) Moderate —— Report only when trigger conditions are met (Default)
• (c) Aggressive —— Report only after completing all tasks

Step 4 · Create Working Directory + TASK.md → Start Immediately

4. TASK.md Template

# {Task Name} · Task Control Document

## 0. Meta Information
- Creation Date: YYYY-MM-DD
- Current Batch: Batch N
- Autonomy Mode: (a/b/c)
- Next Reporting Trigger: {Condition}

## 1. User's Task (Refined Original Words)
- {User's Original Words}

## 2. AI's Understanding of the Task (Version Shown to User for Confirmation)
- Core: ...
- Boundaries: Do X, do not do Y
- Output Format: ...
- User Confirmed: ✅ YYYY-MM-DD

## 3. Task Queue

| # | Task | Status | Dependencies | Expected Output | Tool Suggestions |
|---|------|------|------|----------|----------|
| T1 | ... | ⬜ | - | ... | WebSearch |
| T2 | ... | ⬜ | - | ... | Bash |
| T3 | ... | ⬜ | T1 | ... | Reasoning |
| ... |

## 4. Batch Execution Logs (Reverse Order)

### Batch N - YYYY-MM-DD
**Tasks Executed in This Batch**: T{x}, T{y}, T{z}
**Key Findings**: ...
**Queue Verification Results**:
- Add T{n} (Reason: ...)
- Delete T{m} (Reason: ...)
- Adjust Priority of T{k}
**Trigger Condition**: Not Triggered / Triggered §6.X

## 5. Working Hypotheses (For Research/Evaluation Tasks)
- 🟡 H1: ...
- 🟢 H2: ...

## 6. Conditions to Trigger User Intervention
1. Completion of a milestone batch (deliver important outputs)
2. Premise is falsified
3. True fork decision
4. No new progress in 2 consecutive batches
5. Unexpected major discovery
6. Need user's real data

## 7. Decision Records
- D1 (YYYY-MM-DD): ...

## 8. Final Deliverables
- [ ] FINAL-REPORT.md
- [ ] {Other Deliverables}

5. Execution Loop (6 Steps per Batch, No Skipping)

Start of Batch N:

Step 1 · Select Tasks for This Batch
   - Pick K items from the §3 queue (K is usually 3-8)
   - Prioritize tasks without dependencies and those that can be executed in parallel
   - Mark selected items as 🔄 In Progress

Step 2 · Execute
   - Call any tools as needed:
     · WebSearch / WebFetch / Bash / Read / Write / Edit / Glob / Grep
     · Sub-agent (see §8 for Blindfold Protocol)
     · Any MCP tools
     · Can make parallel calls + run background tasks
   - After completing each task → Write the output into the corresponding sub-file

Step 3 · Verify Output
   - After completing each task → Check if the output meets expectations
   - If not → Mark the issue, may need to redo or decompose into new tasks
   - If yes → Mark as ✅ Completed

Step 4 · Queue Health Check ⚠ Critical
   - Proactively ask yourself 3 questions:
     1. Do the results of this batch reveal tasks that need to be added?
     2. Are there any outdated tasks in the queue (should be deleted / modified)?
     3. Do priorities need to be adjusted?
   - Modify the §3 queue based on the answers
   - This is not passive update, but **active health check**

Step 5 · Write Back to TASK.md
   - Update batch number in §0
   - Update queue status in §3
   - Append Batch log to §4
   - Update hypotheses in §5 (if any)

Step 6 · Self-Check Trigger Conditions
   - If any condition in §6 is met → Stop and report
   - Otherwise → Enter Batch N+1

6. Trigger User Intervention (4+3 Conditions)

1. Completion of a milestone batch —— Deliver important outputs, the user may want to review
2. Premise is falsified —— The fundamental assumption of the task collapses
3. True fork decision —— Involves user values, unique information, or irreversible choices

1. Direction does not converge —— No new progress in 2 consecutive batches
2. Unexpected major discovery —— Evidence that overturns existing conclusions
3. Need user's real data —— Cannot be obtained from public sources

Ritual Question Blacklist (Do NOT Ask)

• ❌ "Am I doing this right" → Make decisions if you have basis
• ❌ "Use Tool A or Tool B" → Choose on your own
• ❌ "Can I do this in the next batch" → Just do it, adjust if wrong
• ❌ "Do you approve of Phase X" → The user can review afterwards
• ❌ "Did I decompose this correctly" → Just run it, adjust if it doesn't work

7. Communication Style Switching Protocol

8. Sub-agent Blindfold Protocol (Required for Evaluation/Blind Tests)

1. Spawn sub-agent (use Agent tool, subagent_type: general-purpose)
2. Sub-agent's task:
   • Search for information / retrieve data
   • Write complete data (including answer/result) into File X; the main agent does not read X
   • Return only the blinded version in the response
3. The sub-agent must have an Honesty Protocol: Never reveal the answer in the response, not even hints
4. After receiving the blinded data, the main agent must first commit the prediction/judgment to File Y (Prohibit reading File X first)
5. Only after recording can you read File X to reveal the answer
6. Compare X and Y, score, and write into the results report

Sub-agent Prompt Template

You are a research assistant sub-agent. The main agent is working on {task} and requires strict blind testing.

Your task:
1. Find N {case types} from {suggested sources}
2. Write the complete data of each case (including {field to be blinded}) into:
   {absolute path}/blind-data-full.md
3. In your response, **only return the blinded version**: {list of fields to keep}
4. **Absolutely do not** include any information about {field to be blinded} in your response

Honesty Protocol:
- Do not use suggestive language ("This finally...", "The result is a bit unexpected")
- If you can't find enough, tell the main agent the actual number, don't fabricate
- Mark uncertain fields as "Not stated in original text"

Briefly report the source and quantity after completion.

9. Position Update Protocol

1. Does the new evidence from this batch change my core judgment?
2. If yes, should I explicitly state "I updated my position" instead of supplementing?
3. Am I maintaining an old position that is no longer supported by new evidence?

10. Eight Failure Modes + Prevention

11. Final Report (FINAL-REPORT.md)

1. Executive Summary (3-5 lines TL;DR)
2. Core Conclusions + Evidence (For research/evaluation tasks, at least 3 independent evidence lines)
3. Counterarguments I Considered (List honestly)
4. Methodological Limitations (Sample size, potential contamination, uncovered angles)
5. Next-Step Suggestions
6. Reusable Methodology (If there are non-obvious findings)

12. Worked Example: Exploration of Liu Yao AI Accuracy

/Users/yunshu/Documents/trae_projects/test/liuyao-exploration/

Task Starting Point

Actual Execution Path

• Batch 1: Research the scope of application and failure conditions of Liu Yao → Found that the user's premise may have hit failure conditions
• Batch 2: Astrological meta-analysis + Barnum effect + opposing perspective → 🔴 Indirect evidence of unpassed gate
• Batch 3: MVP evaluation of 5 case rule reasoning → 80% but with honesty contamination
• Batch 4: Conduct real blind test of 8 cases using sub-agent blindfold protocol → 62.5%
• Final: Four-line cross-validation → Liu Yao is close to random in the user's real scenario

Key Failures and Recovery (Pitfalls Encountered in This Task)

• F1: Asked "Do you approve of Phase 0" → Fixed after user push + written into feedback memory
• F2: Dumped blind test results in table → Fixed after user's "Ah????" + written into feedback memory
• F3: Almost made judgment without ontological measurement → Inserted MVP test after user push
• F4: Did not perform automatic blind test in Round 4 → Inserted real blind test after user push

13. Final Self-Check

• All items in §3 queue are ✅ or explicitly ❌
• FINAL-REPORT.md is completed
• All hypotheses in §5 have final states (no 🟡 pending)
• The user's original task (§1) is clearly answered
• Is it worth沉淀 the methodology of this task into feedback memory (if there are non-obvious findings)