Pre-Task Checklist

Never guess. Never assume. Always verify.

The Problem

Agents forget things when:

• MEMORY.md has stale data (old IPs, wrong paths, expired tokens)
• Context window compaction loses recent details
• Multiple concurrent sub-agents fragment attention
• New tasks override previous context

Result: Wrong IPs, incorrect formats, missed rules, repeated mistakes.

The Solution

Mandatory pre-task checklist before ANY non-trivial task:

1. memory_search() for relevant patterns
2. Read reference files directly (don't assume)
3. Verify critical details in real-time
4. Check AGENTS.md / SOUL.md for rules

Quick Start

# Before starting a task
source ~/.openclaw/workspace/skills/pre-task-checklist/scripts/checklist.sh

# Verify a specific detail (IP, path, format)
verify "GPU server IP"
verify "MBD article format"

# Full checklist (interactive)
run_checklist

Checklist Items

1. Memory Search

# Search for relevant patterns
memory_search "GPU server IP address"
memory_search "MBD article format"
memory_search "folder naming convention"

What to look for:

• Previous examples of this task
• Correct IPs, paths, formats
• Rules or constraints
• Past mistakes to avoid

2. Read Reference Files

# Read actual examples, don't assume
cat /tmp/mbd-book-ideas/2026-03-11_懒人理财经_AI替你管钱躺赚不是梦/文章.md
cat ~/.openclaw/workspace/TOOLS.md
cat ~/.openclaw/workspace/AGENTS.md

What to verify:

• Format and structure
• Tone and style
• Technical requirements
• Examples to follow

3. Verify Critical Details

# Check live data, don't trust memory
ssh peter@10.0.0.30 "echo 'connected'"  # Verify IP works
ls -la /path/to/file                     # Verify path exists
cat ~/.openclaw/workspace/TOOLS.md       # Verify current value

What to verify:

• IPs and hostnames
• File paths and URLs
• API keys and tokens (check expiry)
• Configuration values

4. Check Rules

# Read rule files before acting
cat ~/.openclaw/workspace/AGENTS.md      # Check project rules
cat ~/.openclaw/workspace/SOUL.md        # Check identity/rules
cat ~/.openclaw/workspace/HEARTBEAT.md   # Check recurring tasks

What to check:

• "No pause" rule - execute immediately
• Repo ownership (clawinfra vs AlexChen31337)
• Publish gates (MbD needs explicit approval)
• Safety/security rules

Examples

Example 1: GPU Server Access

❌ Wrong (guessing):

ssh peter@10.0.0.44  # WRONG IP from stale memory

✅ Right (verifying):

# Step 1: Search memory
memory_search "GPU server IP"
# → Returns: "10.0.0.44" (STALE!)

# Step 2: Verify in TOOLS.md
cat ~/.openclaw/workspace/TOOLS.md | grep "GPU Server"
# → Shows: "peter@10.0.0.30" (CURRENT!)

# Step 3: Test connection
ssh peter@10.0.0.30 "hostname"
# → Works!

# Step 4: Update memory
# (Correct the stale data)

Example 2: MBD Article Format

❌ Wrong (assuming):

# Write 8,000-word academic whitepaper

✅ Right (verifying):

# Step 1: Search memory
memory_search "MBD article word count"
# → Returns: "3,000-4,500 words"

# Step 2: Read example
cat /tmp/mbd-book-ideas/2026-03-11_懒人理财经_AI替你管钱躺赚不是梦/文章.md
# → See conversational tone, chapter structure

# Step 3: Check rules
cat AGENTS.md | grep -A5 "MbD publish"
# → "MbD: NEVER publish without explicit 'publish to MbD'"

# Step 4: Follow format
# Write 4,000 words, conversational style

Example 3: GitHub Repo Ownership

❌ Wrong (assuming):

gh repo create clawinfra/my-repo  # WRONG!

✅ Right (verifying):

# Step 1: Check rules
cat AGENTS.md | grep -A10 "GitHub Repo Ownership"
# → "clawinfra/ ONLY for ClawChain, EvoClaw, core infrastructure"
# → "AlexChen31337/ for research, tools, benchmarks"

# Step 2: Classify repo
# My tool is NOT core infrastructure → use AlexChen31337

# Step 3: Create in correct org
gh repo create AlexChen31337/my-repo

Consequences for Violations

When you skip the checklist and make a mistake:

1. Acknowledge immediately

   • "I forgot to verify X"
   • "I assumed Y instead of checking"

2. Write to WAL

   uv run python skills/agent-self-governance/scripts/wal.py append main \
     --type checklist_violation \
     "Failed to verify GPU server IP. Used 10.0.0.44 (stale) instead of 10.0.0.30 (current)."
   

3. Correct before continuing

   • Fix the mistake
   • Update stale data
   • Verify the fix works

4. Identify root cause

   • Why did I skip verification?
   • Was I rushed?
   • Did I feel confident when I shouldn't have?

Integration with Other Skills

agent-motivator

• agent-motivator: "Push harder, don't give up"
• pre-task-checklist: "Verify before you push"

Use together:

1. Run pre-task-checklist to verify the approach
2. Use agent-motivator to push through obstacles

rsi-loop

• rsi-loop: Log outcomes and improve
• pre-task-checklist: Prevent mistakes before they happen

Use together:

1. Run pre-task-checklist before task
2. Log outcome with rsi-loop after task
3. Update checklist based on learnings

skill-creator

When creating new skills, include a "Verification" section that references this checklist.

Verification Scripts

Quick Verify Function

# Add to ~/.bashrc or source directly
verify() {
    local query="$1"
    echo "🔍 Verifying: $query"
    memory_search "$query" | head -5
    echo "📖 Check reference files for: $query"
}

Full Checklist Script

run_checklist() {
    echo "=== PRE-TASK CHECKLIST ==="
    echo "1. What task am I doing?"
    read task
    echo "2. Search memory for: $task"
    memory_search "$task" | head -10
    echo "3. Read reference files? (y/n)"
    read ref_files
    if [ "$ref_files" = "y" ]; then
        echo "Which files?"
        read files
        cat $files
    fi
    echo "4. Verify critical details? (y/n)"
    read verify_details
    if [ "$verify_details" = "y" ]; then
        echo "What to verify?"
        read verify_what
        verify "$verify_what"
    fi
    echo "5. Check rules? (y/n)"
    read check_rules
    if [ "$check_rules" = "y" ]; then
        cat ~/.openclaw/workspace/AGENTS.md
    fi
    echo "=== CHECKLIST COMPLETE ==="
    echo "Proceed with task? (y/n)"
    read proceed
    if [ "$proceed" = "y" ]; then
        echo "✅ Proceeding with verified context"
    else
        echo "❌ Task aborted - need more verification"
    fi
}

Teaching Agents to Use This Skill

When spawning sub-agents, include this in the prompt:

sessions_spawn(
    runtime="subagent",
    task=f"""
    Task: {task_description}
    
    MANDATORY PRE-TASK CHECKLIST:
    1. Search memory for similar tasks
    2. Read reference files (don't assume)
    3. Verify critical details (IPs, paths, formats)
    4. Check rules in AGENTS.md/SOUL.md
    
    Start by running the checklist and reporting what you found.
    Then proceed with the task.
    """
)

Common Pitfalls

Pitfall 1: "I Remember This"

Thought: "I used the GPU server yesterday, I know the IP" Problem: Memory might be stale Fix: Verify in TOOLS.md or test connection

Pitfall 2: "This Is Obvious"

Thought: "Everyone knows MBD articles are 8,000 words" Problem: Assumptions are wrong Fix: Read an actual MBD article example

Pitfall 3: "I'll Check Later"

Thought: "Let me just start, I'll verify the IP later" Problem: You'll forget, and errors will compound Fix: Verify BEFORE starting

Pitfall 4: "Searching Takes Too Long"

Thought: "This will add 2 minutes to every task" Problem: One mistake costs 10+ minutes to fix Fix: 2 minutes now vs 10 minutes later

When to Skip the Checklist

Only for trivial tasks:

• "What time is it?"
• "Send this message"
• "Read this file"
• "List running processes"

Never skip for:

• Using IPs, paths, URLs
• Writing to files
• Publishing content
• Deploying code
• Spawning sub-agents
• Financial operations
• Security-sensitive actions

Success Metrics

Track your improvement:

• Week 1: Checklist on every task (100%)
• Week 2: Checklist on 50% of tasks (build intuition)
• Week 3: Checklist on 20% of tasks (only complex/uncertain)
• Week 4: Checklist becomes automatic habit

Goal: Zero mistakes from stale data or forgotten rules.

---

Remember: The 2 minutes you spend verifying saves 20 minutes fixing mistakes.

When in doubt: CHECK. When confident: CHECK AGAIN.