# Design Shotgun

> Rapid design-variant exploration for Codex. Use when the user wants multiple UI directions, visual options, or a faster feedback loop on taste.

- Skill: `mphaxise/design-shotgun` (Agent Skill)
- Install (CLI): `npx skillmds@latest add mphaxise/design-shotgun`
- Raw SKILL.md: https://api.skillmd.com/api/skills/mphaxise/design-shotgun/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: mphaxise (https://skillmd.com/u/mphaxise)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/mphaxise/design-shotgun

---


# Design Shotgun

Use this skill when the user wants to see several substantially different design directions before committing to one.

This port is adapted from `garrytan/gstack` at commit `4d2c8d94d00cc4f4f3d4c26316a4f939ceedc045`.

## Workflow

1. Capture the product surface, constraints, and non-negotiables.
2. Load `PRODUCT.md`, `DESIGN.md`, or equivalent context plus the existing design system.
3. Produce 4-6 meaningfully different directions, not tiny variations.
4. For each direction, name:
   - the thesis
   - the visual mood
   - the main tradeoff
5. If Impeccable live mode is already installed and the user wants interactive web variants, use it through `design-quality`. Keep its runtime and source-rewrite machinery external.
6. Otherwise generate or assemble visual variants with the available visual tools; use structured descriptions only when visual generation is unavailable.
7. Apply a squint test for hierarchy and a product-context test for distinctiveness.
8. Narrow based on the user's reaction and iterate.

## Guardrails

- Do not waste cycles on superficial color swaps.
- Keep each direction distinct enough that the user can develop taste preferences.
- Stay anchored to the product goal rather than free-associating visually.
- Never install or enable Impeccable implicitly.

## Outputs

Always include:

- the set of design directions
- what makes each one distinct
- recommended favorite and why
- what feedback to collect next
- visual evidence and runtime path used