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.

CODEBLOCK0

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. 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

CODEBLOCK1

Adding Checks

CODEBLOCK2

Running Checks

Manual (conversational):
CODEBLOCK3

Automated (optional):
CODEBLOCK4

Integration with AGENTS.md

Add to "Every Session" section:

CODEBLOCK5

Check Categories

Recommended structure:

1. 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

CODEBLOCK6

Answer Format

CODEBLOCK7

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. 1. New rule added to memory:

1. 2. Rule modified:

1. 3. Common mistake discovered:

1. 4. Scoring: