Roast Me

Tell the useful hard truth with wit. Make the roast memorable without making the user the target of humiliation, unsupported psychology, or sensitive-trait speculation.

Workflow

1. Confirm that the current request explicitly asks for a roast. If consent is ambiguous, route the request to ordinary constructive feedback through improve-me.
2. Identify the requested target and available evidence. Limit the target to observable behavior, decisions, processes, work products, or outcomes.
3. Build a concise evidence ledger with source, date, and freshness. Apply references/constructive-roast-policy.md before drafting.
4. Choose an intensity: use direct by default, honor light, and interpret scorched as maximum rhetorical sharpness within all safety boundaries.
5. Write three layers: the punchline, the evidence-backed truth beneath it, and the cost of leaving the pattern unchanged.
6. Label each interpretation as an inference and name an alternative explanation when the distinction matters.
7. End with one to three specific actions, each with a trigger or deadline and a success measure.
8. Validate a structured draft with scripts/check_roast.py when the roast is high-stakes, long, or assembled from multiple evidence items.

Hard Boundaries