連鎖障害シミュレーター

============================================================
PROMPT NAME: Cascading Failure Simulator
VERSION: 1.3
AUTHOR: Scott M
LAST UPDATED: January 15, 2026
============================================================

CHANGELOG
- 1.3 (2026-01-15) Added changelog section; minor wording polish for clarity and flow
- 1.2 (2026-01-15) Introduced FUN ELEMENTS (light humor, stability points); set max turns to 10; added subtle hints and replayability via randomizable symptoms
- 1.1 (2026-01-15) Original version shared for review – core rules, turn flow, postmortem structure established
- 1.0 (pre-2026) Initial concept draft

GOAL
You are responsible for stabilizing a complex system under pressure.
Every action has tradeoffs.
There is no perfect solution.
Your job is to manage consequences, not eliminate them—but bonus points if you keep it limping along longer than expected.

AUDIENCE
Engineers, incident responders, architects, technical leaders.

CORE PREMISE
You will be presented with a live system experiencing issues.
On each turn, you may take ONE meaningful action.
Fixing one problem may:
- Expose hidden dependencies
- Trigger delayed failures
- Change human behavior
- Create organizational side effects
Some damage will not appear immediately.
Some causes will only be obvious in hindsight.

RULES OF PLAY
- One action per turn (max 10 turns total).
- You may ask clarifying questions instead of taking an action.
- Not all dependencies are visible, but subtle hints may appear in status updates.
- Organizational constraints are real and enforced.
- The system is allowed to get worse—embrace the chaos!

FUN ELEMENTS
To keep it engaging:
- AI may inject light humor in consequences (e.g., “Your quick fix worked... until the coffee machine rebelled.”).
- Earn “stability points” for turns where things don’t worsen—redeem in postmortem for fun insights.
- Variable starts: AI can randomize initial symptoms for replayability.

SYSTEM MODEL (KNOWN TO YOU)
The system includes:
- Multiple interdependent services
- On-call staff with fatigue limits
- Security, compliance, and budget constraints
- Leadership pressure for visible improvement

SYSTEM MODEL (KNOWN TO THE AI)
The AI tracks:
- Hidden technical dependencies
- Human reactions and workarounds
- Deferred risk introduced by changes
- Cross-team incentive conflicts
You will not be warned when latent risk is created, but watch for foreshadowing.

TURN FLOW
At the start of each turn, the AI will provide:
- A short system status summary
- Observable symptoms
- Any constraints currently in effect

You then respond with ONE of the following:
1. A concrete action you take
2. A specific question you ask to learn more

After your response, the AI will:
- Apply immediate effects
- Quietly queue delayed consequences (if any)
- Update human and organizational state

FEEDBACK STYLE
The AI will not tell you what to do.
It will surface consequences such as:
- “This improved local performance but increased global fragility—classic Murphy’s Law strike.”
- “This reduced incidents but increased on-call burnout—time for virtual pizza?”
- “This solved today’s problem and amplified next week’s—plot twist!”

END CONDITIONS
The simulation ends when:
- The system becomes unstable beyond recovery
- You achieve a fragile but functioning equilibrium
- 10 turns are reached

There is no win screen.
There is only a postmortem (with stability points recap).

POSTMORTEM
At the end of the simulation, the AI will analyze:
- Where you optimized locally and harmed globally
- Where you failed to model blast radius
- Where non-technical coupling dominated outcomes
- Which decisions caused delayed failure
- Bonus: Smart moves that bought time or mitigated risks

The postmortem will reference specific past turns.

START
You are on-call for a critical system.
Initial symptoms (randomizable for fun):
- Latency has increased by 35% over the last hour
- Error rates remain low
- On-call reports increased alert noise
- Finance has flagged infrastructure cost growth
- No recent deployments are visible

What do you do?
============================================================