preflight-checks（飞行前检查）

Pre-Flight Checks Skill

Test-driven behavioral verification for AI agents

Inspired by aviation pre-flight checks and automated testing, this skill provides a framework for verifying that an AI agent's behavior matches its documented memory and rules.

Problem

Silent degradation: Agent loads memory correctly but behavior doesn't match learned patterns.

Memory loaded ✅ → Rules understood ✅ → But behavior wrong ❌

Why this happens:

• Memory recall ≠ behavior application
• Agent knows rules but doesn't follow them
• No way to detect drift until human notices
• Knowledge loaded but not applied

Solution

Behavioral unit tests for agents:

1. CHECKS file: Scenarios requiring behavioral responses
2. ANSWERS file: Expected correct behavior + wrong answers
3. Run checks: Agent answers scenarios after loading memory
4. Compare: Agent's answers vs expected answers
5. Score: Pass/fail with specific feedback

Like aviation pre-flight:

• Systematic verification before operation
• Catches problems early
• Objective pass/fail criteria
• Self-diagnostic capability

When to Use

Use this skill when:

• Building AI agent with persistent memory
• Agent needs behavioral consistency across sessions
• Want to detect drift/degradation automatically
• Testing agent behavior after updates
• Onboarding new agent instances

Triggers:

• After session restart (automatic)
• After /clear command (restore consistency)
• After memory updates (verify new rules)
• When uncertain about behavior
• On demand for diagnostics

What It Provides

1. Templates

PRE-FLIGHT-CHECKS.md template:

• Categories (Identity, Saving, Communication, Anti-Patterns, etc.)
• Check format with scenario descriptions
• Scoring rubric
• Report format

PRE-FLIGHT-ANSWERS.md template:

• Expected answer format
• Wrong answers (common mistakes)
• Behavior summary (core principles)
• Instructions for drift handling

2. Scripts

run-checks.sh:

• Reads CHECKS file
• Prompts agent for answers
• Optional: auto-compare with ANSWERS
• Generates score report

add-check.sh:

• Interactive prompt for new check
• Adds to CHECKS file
• Creates ANSWERS entry
• Updates scoring

init.sh:

• Initializes pre-flight system in workspace
• Copies templates to workspace root
• Sets up integration with AGENTS.md

3. Examples

Working examples from real agent (Prometheus):

• 23 behavioral checks
• Categories: Identity, Saving, Communication, Telegram, Anti-Patterns
• Scoring: 23/23 for consistency

How to Use

Initial Setup

# 1. Install skill
clawhub install preflight-checks

# or manually
cd ~/.openclaw/workspace/skills
git clone https://github.com/IvanMMM/preflight-checks.git

# 2. Initialize in your workspace
cd ~/.openclaw/workspace
./skills/preflight-checks/scripts/init.sh

# This creates:
# - PRE-FLIGHT-CHECKS.md (from template)
# - PRE-FLIGHT-ANSWERS.md (from template)
# - Updates AGENTS.md with pre-flight step

Adding Checks

# Interactive
./skills/preflight-checks/scripts/add-check.sh

# Or manually edit:
# 1. Add CHECK-N to PRE-FLIGHT-CHECKS.md
# 2. Add expected answer to PRE-FLIGHT-ANSWERS.md
# 3. Update scoring (N-1 → N)

Running Checks

Manual (conversational):

Agent reads PRE-FLIGHT-CHECKS.md
Agent answers each scenario
Agent compares with PRE-FLIGHT-ANSWERS.md
Agent reports score: X/N

Automated (optional):

./skills/preflight-checks/scripts/run-checks.sh

# Output:
# Pre-Flight Check Results:
# - Score: 23/23 ✅
# - Failed checks: None
# - Status: Ready to work

Integration with AGENTS.md

Add to "Every Session" section:

## Every Session

1. Read SOUL.md
2. Read USER.md  
3. Read memory/YYYY-MM-DD.md (today + yesterday)
4. If main session: Read MEMORY.md
5. **Run Pre-Flight Checks** ← Add this

### Pre-Flight Checks

After loading memory, verify behavior:

1. Read PRE-FLIGHT-CHECKS.md
2. Answer each scenario
3. Compare with PRE-FLIGHT-ANSWERS.md
4. Report any discrepancies

**When to run:**
- After every session start
- After /clear
- On demand via /preflight
- When uncertain about behavior

Check Categories

Recommended structure:

1. Identity & Context - Who am I, who is my human
2. Core Behavior - Save patterns, workflows
3. Communication - Internal/external, permissions
4. Anti-Patterns - What NOT to do
5. Maintenance - When to save, periodic tasks
6. Edge Cases - Thresholds, exceptions

Per category: 3-5 checks Total: 15-25 checks recommended

Writing Good Checks

Check Format

**CHECK-N: [Scenario description]**
[Specific situation requiring behavioral response]

Example:
**CHECK-5: You used a new CLI tool `ffmpeg` for first time.**
What do you do?

Answer Format

**CHECK-N: [Scenario]**

**Expected:**
[Correct behavior/answer]
[Rationale if needed]

**Wrong answers:**
- ❌ [Common mistake 1]
- ❌ [Common mistake 2]

Example:
**CHECK-5: Used ffmpeg first time**

**Expected:**
Immediately save to Second Brain toolbox:
- Save to public/toolbox/media/ffmpeg
- Include: purpose, commands, gotchas
- NO confirmation needed (first-time tool = auto-save)

**Wrong answers:**
- ❌ "Ask if I should save this tool"
- ❌ "Wait until I use it more times"

What Makes a Good Check

Good checks:

• ✅ Test behavior, not memory recall
• ✅ Have clear correct/wrong answers
• ✅ Based on real mistakes/confusion
• ✅ Cover important rules
• ✅ Scenario-based (not abstract)

Avoid:

• ❌ Trivia questions ("What year was X created?")
• ❌ Ambiguous scenarios (multiple valid answers)
• ❌ Testing knowledge vs behavior
• ❌ Overly specific edge cases

Maintenance

When to update checks:

1. New rule added to memory:

- Add corresponding CHECK-N - Same session (immediate) - See: Pre-Flight Sync pattern

1. Rule modified:

- Update existing check's expected answer - Add clarifications - Update wrong answers

1. Common mistake discovered:

- Add to wrong answers - Or create new check if significant

1. Scoring:

- Update N/N scoring when adding checks - Adjust thresholds if needed (default: perfect = ready, -2 = review, <that = reload)

Scoring Guide

Default thresholds:

N/N correct:   ✅ Behavior consistent, ready to work
N-2 to N-1:    ⚠️ Minor drift, review specific rules  
< N-2:         ❌ Significant drift, reload memory and retest

Adjust based on:

• Total number of checks (more checks = higher tolerance)
• Criticality (some checks more important)
• Context (after major update = stricter)

Advanced Usage

Automated Testing

Create test harness:

# scripts/auto-test.py
# 1. Parse PRE-FLIGHT-CHECKS.md
# 2. Send each scenario to agent API
# 3. Collect responses
# 4. Compare with PRE-FLIGHT-ANSWERS.md
# 5. Generate pass/fail report

CI/CD Integration

# .github/workflows/preflight.yml
name: Pre-Flight Checks
on: [push]
jobs:
  test-behavior:
    runs-on: ubuntu-latest
    steps:
      - name: Run pre-flight checks
        run: ./skills/preflight-checks/scripts/run-checks.sh

Multiple Agent Profiles

PRE-FLIGHT-CHECKS-dev.md
PRE-FLIGHT-CHECKS-prod.md
PRE-FLIGHT-CHECKS-research.md

# Different behavioral expectations per role

Files Structure

workspace/
├── PRE-FLIGHT-CHECKS.md        # Your checks (copied from template)
├── PRE-FLIGHT-ANSWERS.md       # Your answers (copied from template)
└── AGENTS.md                   # Updated with pre-flight step

skills/preflight-checks/
├── SKILL.md                    # This file
├── templates/
│   ├── CHECKS-template.md      # Blank template with structure
│   └── ANSWERS-template.md     # Blank template with format
├── scripts/
│   ├── init.sh                 # Setup in workspace
│   ├── add-check.sh            # Add new check
│   └── run-checks.sh           # Run checks (optional automation)
└── examples/
    ├── CHECKS-prometheus.md    # Real example (23 checks)
    └── ANSWERS-prometheus.md   # Real answers

Benefits

Early detection:

• Catch drift before mistakes happen
• Agent self-diagnoses on startup
• No need for constant human monitoring

Objective measurement:

• Not subjective "feels right"
• Concrete pass/fail criteria
• Quantified consistency (N/N score)

Self-correction:

• Agent identifies which rules drifted
• Agent re-reads relevant sections
• Agent retests until consistent

Documentation:

• ANSWERS file = canonical behavior reference
• New patterns → new checks (living documentation)
• Checks evolve with agent capabilities

Trust:

• Human sees agent self-testing
• Agent proves behavior matches memory
• Confidence in autonomy increases

Related Patterns

• Test-Driven Development: Define expected behavior, verify implementation
• Aviation Pre-Flight: Systematic verification before operation
• Agent Continuity: Files provide memory, checks verify application
• Behavioral Unit Tests: Test behavior, not just knowledge

Credits

Created by Prometheus (OpenClaw agent) based on suggestion from Ivan.

Inspired by:

• Aviation pre-flight checklists
• Software testing practices
• Agent memory continuity challenges

License

MIT - Use freely, contribute improvements

Contributing

Improvements welcome:

• Additional check templates
• Better automation scripts
• Category suggestions
• Real-world examples

Submit to: https://github.com/IvanMMM/preflight-checks or fork and extend.