Crucible

Stop wasting time on "analysis paralysis" or "gut feelings" for high-stakes decisions. This skill transforms your agent into a structured council that stress-tests ideas through adversarial debate. It moves beyond simple chat by spawning multiple specialized personas—like the Contrarian, First Principles, and Operator—to tear down and rebuild your strategy before you commit resources.

What it does

• Multi-persona debate spawns up to 17 specialized sub-agents to analyze your idea from conflicting angles.
• Structured duels pit personas against each other in anonymized rebuttals to ensure arguments win on merit, not tone.
• Automated model routing utilizes Opus for the chair and critical thinkers while offloading rubric checks to faster models.
• Decision logging writes full, high-fidelity transcripts to disk, keeping your main chat clean while preserving the reasoning.
• Intake guardrails audit your motivations to prevent the crucible from becoming a procrastination ritual.

How it works

1. Intake triggers the chair to classify your input as a problem or solution and restate the question for confirmation.
2. Opening takes run parallel sub-agents to provide independent, non-contaminated analysis based on strict rubrics.
3. Pair duels execute serial rebuttals where personas attack each other's arguments without knowing who wrote them.
4. Verdict delivers a final Proceed, Reframe, or Kill recommendation alongside a mandatory minority report.

Frameworks & tools

The skill is designed for advanced LLM environments supporting multi-agent orchestration. It utilizes specific model routing across different tiers and requires local file system access to manage the crucible-transcripts/ directory.

Why this beats prompting it yourself

Standard prompting often leads to "mediator mush" where the AI tries to be too helpful or agreeable. This skill enforces adversarial tension and strict rubrics, ensuring you hear the dissenting voices that single-prompt sessions usually bury.

Use cases

• Strategic planning for side projects or product directions where the cost of being wrong is high.
• Sanity checks for daily PM ideation to quickly identify hidden flaws in a feature request.
• Career moves that require a heavy-lift "existential" council to evaluate long-term consequences.
• Framework reframing when you suspect you are pattern-matching correctly but asking the wrong question.

Known limitations

High resource usage requires 30-90 seconds for full councils. Requires a local directory for transcript persistence. Not intended for tactical, low-stakes questions.