Roast-me: a skill to teach LLMs to be harsh

A new Claude Code skill called Roast-me subjects raw ideas to three adversarial rounds of questioning from personas including a skeptical user, a numerate investor, and a shipping engineer, then delivers a verdict and optional pitch or implementation plan. The skill, which lives in .claude/skills/roast-me/, forces ideas into a one-sentence format and checks against named failure modes such as 'vitamin not painkiller' and 'distribution unsolved'.

A Claude Code skill that interrogates a raw idea instead of encouraging it, plus the archive of ideas that have been through it. The premise: every question the roast asks is one you will get asked later by someone whose money or time is on the line. Better to hear it here. You drop an idea. The skill runs it through three adversarial rounds, each with a different persona, then writes a verdict and, if anything survived, a pitch or an implementation plan. Phase 0, capture. Forces the idea into one sentence: for who , what , instead of current alternative . Creates ideas/<slug /idea.md with the one-liner, your raw description verbatim, and the date. Phase 1, recon. Quietly finds who already does this and whether there's evidence the problem is real, then folds the strongest competitor into the first round of questions. Phase 2, three rounds. Round 1 is the target user who doesn't care and already gets by. Round 2 is a numerate investor who has seen this before. Round 3 is the engineer who has to ship it and the operator who has to run it. Max four questions per round. Phase 3, verdict. Writes ideas/<slug /roast.md : kill shots, what survived, everything you hand-waved, the single riskiest assumption plus the cheapest test for it, and a plain BUILD / RESHAPE / PARK . Phase 4, deliverable. A one-page pitch, a risk-ordered implementation plan, or both, built only from what survived. Ground rules the skill holds itself to: no flattery, no compliment sandwiches, attacks on the idea and never the person, every attack specific and falsifiable, vague answers marked UNANSWERED rather than quietly filled in with optimism. From this directory, in Claude Code: /roast-me Or just say "roast this idea", "poke holes in this", or "would you invest in this" and describe the idea. The skill triggers on its own. Answer the questions honestly. A roast you win by being vague produces a pitch that fails on contact with a real audience, which defeats the point. .claude/skills/roast-me/ SKILL.md the roast itself: rules, personas, phases references/ attack-lines.md question bank + named failure modes pitch-template.md one-page pitch structure plan-template.md risk-ordered plan structure ideas/ <slug / idea.md one-liner + raw description + date roast.md kill shots, survivors, verdict pitch.md optional deliverable plan.md optional deliverable references/attack-lines.md carries the named failure modes the roast checks against: vitamin not painkiller , feature not a company , founder is the only user , distribution unsolved , economics inverted , and a dozen more. It's a bank to draw from, not a checklist to run. | Idea | Status | |---|---| slm-agentic-ops | agentops-energy mixed-ai-node-case