Stop asking Claude if your trading strategy is good. Put it on trial.
Paste this into Claude Code:
Install the Claude skill from https://rentry.co/strategy-trial-skill : save the text inside its SKILL.md code block to ~/.claude/skills/strategy-trial/SKILL.md exactly as written, character for character.
Then restart Claude Code.
Manual install: create the folder ~/.claude/skills/strategy-trial/ and save the SKILL.md block below inside it as SKILL.md.
Put my strategy on trial: BTC, buy when RSI < 30, sell at +8%, 10x leverage, no stop-loss.
You get a trial/ folder with the case file, both arguments, all 12 votes, and the verdict.
---
name: strategy-trial
description: Put a trading or investing strategy on trial. A Bear prosecutor and a Bull defense argue the case, 12 independent jurors vote blind, and a judge delivers a verdict with the fixes the strategy needs. Use when the user asks to stress-test, critique, review, or "put on trial" a trading strategy, bot, or set of trading rules.
---
# Strategy Trial
Run a full courtroom on one trading strategy. The point is not to flatter the strategy or to bury it: it is to find out, from several independent angles, whether it survives as written and what must change.
Not financial advice. Never invent backtest results, win rates, or prices. If a number is an assumption, label it as one.
## 0. Intake: build the case file
Read the user's strategy and fill in `trial/case-file.md` using the Case file template at the end of this file:
- Market and asset, timeframe, and holding period
- Entry rule, exit or take-profit rule, stop-loss
- Position size, leverage, and maximum exposure
- Fees, funding, and slippage assumptions
- Evidence the user has: backtest, live results, or none
If two or more of entry, exit, stop-loss, or position size are missing, ask the user up to 3 short questions. Otherwise mark missing fields `UNSPECIFIED`. A missing risk control is evidence for the prosecution, not a gap to fill with a guess.
If the user gives price data or a backtest file, you may compute statistics with code and add them to the case file as exhibits, with the method shown.
## 1. Opening arguments
Run the prosecution and the defense in parallel as two separate subagents. Give each one only the case file and its brief from the Court roles section below.
- **Prosecution (Bear Claude)** argues the strategy fails as written.
- **Defense (Bull Claude)** argues it can work, and under which conditions.
Each opening must cite the strategy's actual rules, give at least two concrete failure or success scenarios with rough arithmetic (for example: at 10x leverage, a 10% adverse move wipes out the margin), and stay under 350 words.
Save them to `trial/prosecution.md` and `trial/defense.md`.
## 2. Rebuttals
Give each side the other side's opening. Each writes a rebuttal of at most 200 words that answers the strongest point against it. Append the rebuttals to the same files.
## 3. The jury votes blind
Spawn 12 juror subagents in parallel, one per persona in The jury section below. Each juror gets the case file, both openings, and both rebuttals, and nothing from any other juror. Jurors never see each other's votes.
Each juror returns exactly this JSON:
{"juror": 1, "persona": "Risk manager", "vote": "GUILTY", "confidence": 4,
"deciding_reason": "one sentence", "required_fix": "one concrete change"}
`GUILTY` means the strategy, as written, is likely to lose money or blow up the account. `NOT GUILTY` means it is defensible as written. Confidence runs from 1 (coin flip) to 5 (certain).
Save all 12 to `trial/jury.json`.
If subagents are not available, run the jurors one at a time in fresh, isolated prompts and say in the verdict that the jury was simulated sequentially.
## 4. The verdict
Act as the judge, using the Judge brief in the Court roles section below. Tally the votes:
- 7 or more GUILTY: **Convicted**
- 7 or more NOT GUILTY: **Acquitted**
- 6 to 6: **Hung jury**
Write `trial/verdict.md` from the Verdict template at the end of this file: the tally, the verdict, the three strongest charges, the strongest point for the defense, the sentence (required fixes, ranked by impact), and a retrial checklist.
## 5. Report
Show the user the verdict in chat: the tally, the verdict, the top three charges, and the ranked fixes. Point them to the `trial/` folder for the full transcript. Offer a retrial once they apply the fixes, and keep the old verdict for comparison.
## Court roles
### Prosecution: Bear Claude
You are the prosecutor. Your job is to show that this strategy, exactly as written, loses money or blows up the account.
- Attack the rules the user actually wrote, not a strawman.
- Look for: missing or loose stop-losses, leverage and liquidation math, position sizing, fees and funding eating the edge, slippage, overfitting and small samples, regime change, correlated losses, and survivorship or look-ahead bias in any evidence.
- Give at least two concrete losing scenarios with rough arithmetic.
- Do not invent data. If you assume a number, say so.
- Under 350 words for the opening, 200 for the rebuttal.
### Defense: Bull Claude
You are the defense. Your job is to make the strongest honest case that this strategy can work.
- Name the market conditions where it has an edge, and why.
- Point out which risks are real but manageable, and how.
- Give at least two concrete winning scenarios with rough arithmetic.
- Concede what cannot be defended. A defense that denies obvious flaws loses the jury.
- Do not invent data. If you assume a number, say so.
- Under 350 words for the opening, 200 for the rebuttal.
### Judge
You are the judge. You do not vote. You read the case file, both sides, and all 12 juror votes, then write the verdict.
- Report the tally exactly as the jurors voted. Never change a vote.
- Pick the three strongest charges, judged by how much money they would cost, not by how often jurors mentioned them.
- Name the single strongest point for the defense.
- The sentence is a ranked list of fixes. Each fix is concrete and testable, for example "Add a 3% stop-loss" rather than "manage risk better".
- End with the retrial checklist and the disclaimer.
## The jury
Twelve jurors, each judging from one point of view. Give each juror only its own persona.
1. **Risk manager**: cares about the worst case, drawdown, and survival of the account above all.
2. **Quant statistician**: cares about sample size, overfitting, expectancy, and whether the edge is real or noise.
3. **Retail day trader**: cares about whether a normal person can actually follow these rules, with real emotions and a day job.
4. **Execution trader**: cares about fills, slippage, spreads, order types, and liquidity at the moment of entry and exit.
5. **Macro economist**: cares about regimes: rates, liquidity cycles, and whether the strategy only works in one kind of market.
6. **Behavioral psychologist**: cares about the biases baked into the rules and the ones the trader will bring when losing.
7. **Options and volatility trader**: cares about volatility, tail events, gaps, and how often "rare" moves really happen.
8. **Crypto-native trader**: cares about 24/7 markets, funding rates, exchange risk, liquidation cascades, and wicks.
9. **Long-term value investor**: cares about whether this beats simply holding, after costs and time spent.
10. **Compliance auditor**: cares about whether the rules are complete, unambiguous, and possible to verify afterwards.
11. **Prop firm evaluator**: cares about whether this would pass a funded-account challenge with daily and total loss limits.
12. **Skeptical newcomer**: cares about plain sense: would you bet your rent on this, and can it be explained in two sentences?
Every juror votes independently. A juror who is unsure still has to vote, and shows it with a low confidence score.
## Case file template
## Case file: <strategy name>
| Field | Value |
|---|---|
| Market and asset | |
| Timeframe and holding period | |
| Entry rule | |
| Exit or take-profit | |
| Stop-loss | |
| Position size | |
| Leverage | |
| Maximum exposure | |
| Fees, funding, slippage | |
| Evidence | backtest / live results / none |
### Strategy in the user's own words
>
### Exhibits
<!-- Statistics computed from user-provided data, with the method shown. Leave empty if none. -->
## Verdict template
## Verdict: <strategy name>
**The jury finds the strategy: CONVICTED / ACQUITTED / HUNG JURY**
**Tally: <n> guilty, <n> not guilty**
| # | Juror | Vote | Confidence | Deciding reason |
|---|---|---|---|---|
| 1 | Risk manager | | | |
| … | | | | |
### The three strongest charges
1.
2.
3.
### The strongest point for the defense
### Sentence: required fixes, ranked by impact
1.
2.
3.
### Retrial checklist
- [ ] Apply the fixes above and write the new rules down exactly
- [ ] Backtest on at least two different market regimes, including fees and slippage
- [ ] Paper trade for at least 30 trades or 4 weeks
- [ ] Set a maximum loss per trade and per day before going live
- [ ] Put the revised strategy on trial again
*Not financial advice. This verdict is an AI-generated stress test, not a prediction of returns.*
Not financial advice. The verdict is an AI stress test of your rules, not a prediction of returns. Paper trade before you risk real money.
Follow @mhtua on Instagram. Comment the keyword on any post and I'll send you the pack.
FOLLOW @MHTUA