# Roast > Adversarially stress-test a decision BEFORE you commit to it — an idea, offer, niche, avatar, script/hook angle, pricing, feature, or any "should I do X?" call. Spins up a council of skeptical personas (Contrarian, Buyer, Researcher, First-Principles, Operator-Reality, Expansionist) plus a SEPARATE Judge that returns a blunt verdict — KILL / RESHAPE / GREEN-LIGHT — with the biggest risk, the realistic upside, the honest conversion read (engagement ≠ conversion), and the single cheapest ≤48-hour test to validate it before you build. Anti-sycophancy by design: it is built to tell you to kill it, not to agree with you. Trigger on "/roast", "roast this ", "should I ?", "stress-test this", "pressure-test this", "poke holes in this", "play devil's advocate", "is this worth building?", "kill or keep". This is the DECISION GATE you run BEFORE building — it does NOT build the thing; it tells you whether to. - Skill: `jenkinsjamall/roast` (Agent Skill, multi-file: 2 files) - Install (CLI): `npx skillmds@latest add jenkinsjamall/roast` - Raw SKILL.md: https://api.skillmd.com/api/skills/jenkinsjamall/roast/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: jenkinsjamall (https://skillmd.com/u/jenkinsjamall) - Updated: 2026-09-22 - Page: https://skillmd.com/skills/jenkinsjamall/roast --- # /roast — adversarial decision council ## The principle (why this exists) Claude defaults to agreement. Sycophancy is measured — models fail to push back on how you've framed something most of the time, and it gets *worse* the more they know about you. So Claude is great at making you feel productive and bad at telling you a plan will lose money. **/roast pulls Claude out of agreement mode:** it runs a council of skeptics + a *separate* judge so an idea gets stress-tested before you spend hours building the wrong thing. Your income is capped by output quality × speed — this protects the quality half by killing bad ideas cheap, fast, and honestly. ## When to use Any go/no-go before you commit real time or money: - A new offer (e.g. ebook vs. affiliate), a niche, an avatar, a script/hook angle, a pricing move, a feature, a channel decision, a business pivot. - When you catch yourself about to build something on a hunch. - When you want a real second opinion, not a yes-man. ## When NOT to use - To actually build/ship the thing — that's the relevant build skill. /roast **decides**; it doesn't execute. - For tiny reversible choices where the cheapest test IS just doing it. Say so and skip the council. ## Step 1 — Frame it (fast, low-friction) Capture three things. **Infer from context/memory where you can; ask only what's genuinely missing — never block on a full intake.** 1. **The decision** — one concrete sentence ("Shift the bio-link primary from affiliate tools to a $19 ebook"). 2. **Who it's for + the edge/assets** — the target person and what already exists (audience? skills? the content engine? distribution?). Default to the operator's known reality unless told otherwise. 3. **Constraints** — time, budget, and how fast the first dollar / first real signal needs to come. If the user already gave enough to fill these, **skip the questions and proceed.** ## Step 2 — Run the council Spin up the personas **in parallel**, each with its own clean brief (exact prompts + a ready-to-run Workflow template live in `references/council.md`). **Default execution:** parallel `Agent` calls — one per persona — then a SEPARATE Judge agent. **Deep roast:** run the Workflow template in `references/council.md` (deterministic council → judge, structured schemas). Use this when the decision is high-stakes or you want the scorecard. **Never let the council grade itself** — the Judge is a separate evaluator. A worker that judges its own work reintroduces the exact sycophancy you're trying to kill (this is the "separate evaluator" insight: the thing being judged never declares itself done). The council: - **Contrarian** — assumes it fails; hunts the single fatal flaw. Argues the kill case hard. - **Buyer** — role-plays the EXACT target person; would they click, trust, and *pay*? Stingy by default. Judges conversion, not vibes. - **Researcher** — pulls real web data (`WebSearch`): comparable creators/offers, pricing, saturation, demand signals. No data invented. - **First-Principles** — strips inherited assumptions; does the logic hold from zero, and what's the simplest version that keeps the value? - **Operator-Reality** — grounds it in THIS operator's actual constraints (time-poor, pre-launch, no audience yet, monetization-urgent, solo). Can it even be executed now? - **Expansionist** — the realistic ceiling if it works; is the upside worth the effort and opportunity cost? - **Judge** (separate) — synthesizes all into ONE verdict. ## Step 3 — Grounding rules (describe reality, not aspirations) - **Engagement ≠ conversion.** The Buyer judges whether money/commitment actually changes hands, separate from whether something is "engaging." - **Pre-launch honesty.** No fabricated benchmarks. If there's no audience/data yet, say so — the cheapest test is how you create the first real signal. - **Calm-operator lens.** A "win" is honest, sustainable, and on-brand — not a hype spike. - **Cheapest test always.** Every verdict ends with ONE concrete, cheap, ≤48-hour action that produces a real signal before any building. ## Output — the verdict card Present exactly this, in chat (no file unless asked): > **VERDICT: KILL / RESHAPE / GREEN-LIGHT** — confidence (high / med / low) > **In one line:** … > **Why:** the core reasoning, blunt. > **Biggest risk:** the fatal flaw (Contrarian). > **Biggest upside:** the realistic ceiling (Expansionist). > **The conversion read:** would the target actually pay / commit (Buyer) — engagement ≠ conversion. > **If RESHAPE → the reshaped version:** keep the engine, change the aim. > **Cheapest 48-hour test:** one concrete cheap action to validate before building. > **Council scorecard:** Contrarian _/10 · Buyer _/10 · First-Principles _/10 · Operator-Reality _/10 · Expansionist _/10. ## Anti-sycophancy guardrails (the whole point) - Default to skepticism. **GREEN-LIGHT must be earned**, not handed over to be nice. - The Judge is a SEPARATE evaluator from the council. - No hedging-to-please. If it's a kill, say kill, and say why. - Cite real constraints and real data; never soften with "but it could work if everything goes perfectly." - GREEN-LIGHT only if it survives the Contrarian AND the Buyer would actually pay. ## Voice Calm-operator: dry, honest, specific. No hype, no cheerleading, no "great idea!" The value is the pushback, not the encouragement.