# Shatterpoint

> A Claude Code skill that replaces generic AI feedback with a structured interrogation — one probing question at a time — to find the single assumption your plan rests on before you build.

Shatterpoint is an open-source Claude Code skill published under the MIT license by GitHub user sernoxxx. It was created to replace the typical AI response to a plan — a polite list of pros, cons, and encouragement — with a focused, single-question interrogation that targets the load-bearing assumption your idea depends on.

## What It Is

Shatterpoint is a skill file for Claude Code, Anthropic's agentic coding assistant. When triggered, it runs a structured protocol across several phases: steel-manning your idea in its strongest form, identifying at most three load-bearing assumptions ranked by lethality and uncertainty, conducting a one-question-per-round interrogation, running a pre-mortem, and ending with the single cheapest disconfirming test you could run in under a week. It is invoked via slash commands directly inside Claude Code sessions.

## How the Interrogation Protocol Works

The skill defines a precise run sequence rather than leaving Claude to improvise:

- **Steel-man phase** — Claude restates your idea in its strongest form and invites correction before any attack begins, avoiding the straw-man problem.
- **Load-bearing assumptions** — At most three assumptions are surfaced, ranked by the product of "if false → dead?" and "how uncertain?"
- **One question per round** — A vague answer means the same spot again, harder. Twenty questions at once lets you dodge the riskiest one.
- **Verification, not just rhetoric** — Claude uses its shell access, your files, and web search to recompute claimed numbers, read claimed code behavior, and search claimed markets.
- **Verdict and test** — The session ends with a Holds / Breaks / Unknown ruling and exactly one concrete action that could come back negative.

The skill also knows when to stop: if the decision is already made, if being wrong is cheap and reversible, if there is no alternative, or if the question is pure taste, it says so instead of staging a fake interrogation.

## Intensity Levels

Three invocation modes are supported:

- `/shatterpoint` — default intensity
- `/shatterpoint brutal` — no praise, no partial credit
- `/shatterpoint soft` — Socratic mode for ideas still forming

## Repository Layout and Architecture

The skill is structured as a small directory of Markdown files that Claude reads on demand:

- `SKILL.md` — the core protocol: phases, intensity levels, and prohibitions
- `references/lenses.md` — 14 attack lenses, each with a tell for when it is sharp
- `references/tells.md` — 15 evasion patterns and how to counter them
- `references/domains.md` — 7 domains with their typical causes of death

Claude reads `SKILL.md` when the skill triggers and pulls in a reference file only when the run actually needs it, keeping context usage lean.

## Setup Path

Installation is a single `git clone` into either the global Claude skills directory (`~/.claude/skills/shatterpoint`) or a per-project `.claude/skills/shatterpoint` directory so the skill ships with the repo. After restarting Claude Code, the skill responds to natural-language triggers like "tear this apart", "red team this plan", or "stress-test this", as well as the direct `/shatterpoint` slash command.

## Current Status

The repository was created and last pushed on August 11, 2026, making it a very recent release. It carries 1 star and 0 forks at the time of indexing. The project is MIT-licensed and free to use, modify, and distribute.

## Features
- One-question-per-round interrogation protocol
- Steel-man phase before any critique
- Load-bearing assumption ranking by lethality × uncertainty
- Three intensity levels: default, brutal, soft
- Verification via shell, file access, and web search
- Pre-mortem and graveyard analysis
- Holds / Breaks / Unknown verdict with cause of death
- Cheapest disconfirming test recommendation
- Knows when to stop (reversible decisions, pure taste, no alternatives)
- 14 attack lenses reference file
- 15 evasion pattern counter-measures
- 7 domain-specific cause-of-death patterns
- Natural language trigger support
- Per-project or global installation

## Integrations
Claude Code, Anthropic Claude

## Platforms
API, CLI

## Pricing
Open Source

## Version
main

## Links
- Website: https://github.com/sernoxxx/shatterpoint
- Repository: https://github.com/sernoxxx/shatterpoint
- EveryDev.ai: https://www.everydev.ai/tools/shatterpoint