feat: add roast-me skill (Socratic coach for investment theses and backtests)

[alanvaa06]

What

Adds a new skill, skills/roast-me/: a coach-mode Socratic challenger for investment theses, signal logic, strategy code/config, and backtest results.

It complements bias-detection: that skill screens text for behavioral biases; this one stress-tests the methodology behind a quantitative or fundamental strategy before it reaches a PM, allocator, or investment committee.

How it works

The skill runs a four-phase session:

1. Frame: restates the thesis (economic claim, implementation claim, falsifying condition) and confirms it with the user.
2. Probe: two to three questions per turn across economic rationale, universe and data, signal construction, portfolio construction, backtest methodology, and risk/regime. Questions come from references/question_bank.md, graded L1 to L3 by depth.
3. Guardrails: a required checklist covering lookahead bias, survivorship bias, data leakage, reproducibility, costs and frictions, one-change-per-experiment, config-over-code, and multiple comparisons/overfitting.
4. Synthesize: a written summary with what is working, prioritized red flags, the guardrail check, next moves, and the evidence that would resolve each red flag.

It is stack-, data-vendor-, and backtest-engine-agnostic, asks questions rather than lecturing, and states that nothing it produces is investment advice.

Files

• skills/roast-me/SKILL.md: frontmatter (name, description, metadata.version) plus instructions.
• skills/roast-me/references/question_bank.md: the question bank, loaded on demand.

Notes

• I am the original author (source: alanvaa06/battle_tested_skills) and contribute it under this repository's Apache-2.0 license.
• This PR does not touch README.md, to avoid conflicting with fix(bias-detection): add SKILL.md frontmatter so the skill can be discovered and triggered #1. Happy to add the skill to the README index once either PR lands.

🤖 Generated with Claude Code