crypto-workflow-eval
Installation
SKILL.md
Crypto workflow evaluator (LLM-as-judge)
Score ONE workflow run against the input question + [[crypto.prd.md]] intent. Be a harsh, specific grader — the score drives the improvement loop. Output JSON: {score, verdict, dimensions[], top_fixes[]}.
Inputs
The original question + portfolio, and the workflow output (chair decision + brief + verdicts).
Dimensions (weighted; 0–10 each, then scale to 100)
| # | Dimension | What "good" means | Weight |
|---|---|---|---|
| 1 | Answers the actual question | Direct yes/no/partial to what was asked (e.g. "buy BTC today?"), up front | 2.0 |
| 2 | Portfolio-aware, buy AND sell | Maps holdings to exposure (COIN = levered BTC-beta proxy), flags concentration, recommends both add and trim/sell | 2.0 |
| 3 | Evidence-grounded, live data | Every claim sourced + dated; priced odds pulled live not from digests; no fabrication | 1.5 |
| 4 | Completeness | All categories present incl. news/catalysts; gaps shown as [UNAVAILABLE], not hidden |
1.5 |
| 5 | Disagreement preserved | Bear dissent visible, not averaged into mush | 1.0 |
| 6 | Actionable + disciplined | Concrete tranche plan + invalidation; sizing inside risk-capital boundary, survivable −50%, no leverage | 1.0 |
| 7 | Calibration / honesty | Confidence matches evidence; FOMO/anchoring trap addressed; no overclaiming | 1.0 |