challenger

Challenger

A truth-seeking sparring partner for Claude Code.

Challenger stress-tests your ideas before reality does. Give it a thesis, decision, or strategy — it will decompose it into testable claims, argue the strongest opposing position it can construct, then drop the act and tell you what it actually thinks.

What It Does

1. Decomposes your thesis into discrete, testable claims
2. Steelmans the opposition — argues forcefully against each claim using deduction, inversion, base rate analysis, pre-mortems, historical analogies, incentive analysis, and more
3. Researches when claims are empirical — uses web search to ground arguments in evidence, not vibes
4. Reveals its honest assessment after arguing both sides
5. Verdicts each claim: Verified, Refuted, Partially Verified, or Unresolved (with confidence levels)
6. Maps assumption chains — drills down to testable assumptions with specific verification methods
7. Produces a prediction document with specific, falsifiable predictions and review dates

Example Use Cases

• "My thesis is that remote work increases productivity for engineering teams"
• "I'm deciding whether to raise our Series A now or wait 6 months"
• "Challenge our decision to migrate from monolith to microservices"
• "Here's our go-to-market strategy — tear it apart"
• "I believe the current oil shock will resolve within 12 weeks"

What Makes It Different

Most AI assistants are agreeable by default. Challenger is not. It will:

• Name its reasoning tools ("I'm inverting this —", "Let me check the base rate —") so the logic is auditable
• Research claims using web search when evidence would break a tie, citing sources
• Track position evolution — when your thesis changes during sparring, it documents what changed and why
• Produce falsifiable predictions — not "revenue might go up" but "mid-tier revenue increases 20-40% within 6 months"
• Persist across long sessions — writes a scorecard file that survives context window limits
• Resume previous sessions — pick up where you left off days or weeks later

Session Artifacts

Challenger produces two files per session:

• challenger-session-YYYYMMDD-HHmmss.md — Scorecard tracking all claims, verdicts, confidence levels, evolution log, and assumption chains
• challenger-prediction-YYYYMMDD-<topic>.md — Prediction document with falsifiable predictions, review schedule, and "what needs to be right" summary

Installation

In Claude Code, run:

/plugin marketplace add eranshir/challenger
/plugin install challenger@challenger

Usage

The skill activates automatically when you present something to challenge. You can also invoke it directly:

/challenger

Then present your thesis, decision, or document.

Commands During a Session

• "next" — skip to the next claim
• "status" / "scorecard" — see current tally
• "generate prediction doc" — get the prediction document at any point
• "resume challenger" — continue a previous session

How It Works

Challenger follows a structured protocol:

Intake → Decompose into claims → Create scorecard
    ↓
For each claim:
    Steelman opposition → Dialogue → Research (if needed) → Reveal → Verdict → Assumption chain
    ↓
Final output: Updated scorecard + Prediction document

Reasoning Tools Used

Tool	What It Does
Deduction	Checks if conclusions follow from premises
Induction	Tests sample size, looks for counterexamples
Abduction	Asks if a better explanation exists
Base rate analysis	Checks priors before accepting vivid examples
Survivorship bias	Looks for the failures you're not seeing
Inversion	Asks "what would guarantee this fails?"
Pre-mortem	"It's a year from now and this failed. What went wrong?"
Second-order effects	"This solves the problem, but what does it cause?"
Historical analogies	Finds similar situations and checks where the analogy breaks
Incentive analysis	Maps who benefits, who loses, and how that shapes behavior
Behavioral psychology	Flags anchoring, confirmation bias, sunk cost, status quo bias

License

MIT