In the Skill's Own Terms
design-shotgun is a visual design exploration tool that generates multiple AI design variants in parallel, opens them side-by-side in a comparison board served over HTTP, and collects structured feedback. It remembers the user's taste across sessions via a persistent taste profile (fonts, colors, layouts, aesthetics) and biases new generations toward demonstrated preferences.
The skill enforces anti-convergence (every variant must use different fonts, palettes, and layouts) and saves all artifacts to ~/.gstack/projects/$SLUG/designs/ for cross-session persistence. It supports both new explorations and revisiting prior sessions, with feedback loops for regeneration, remixing elements across variants, and iteration until the user approves a direction.
What it produces
- N variant PNGs saved to ~/.gstack/projects/$SLUG/designs/<screen-name>-YYYYMMDD/variant-{letter}.png
- HTML comparison board served over HTTP at random port with feedback collection UI (design-board.html)
- approved.json file recording preferred variant, ratings, comments, and date for cross-session taste learning
- Updated taste-profile.json (schema v1) with confidence-weighted approved/rejected patterns per dimension
How It Works
- 01Session Detection and Taste Recovery
Checks ~/.gstack/projects/$SLUG/designs/ for approved.json files from prior sessions and reads the persistent taste-profile.json (schema v1 with confidence decay 5% per week).
- 02Parallel Variant Generation with anti-convergence
Launches N Agent subagents (one per variant) in a single message for concurrent execution.
- 03HTTP Comparison Board with Feedback Loop
Runs $D compare --serve to generate HTML board and start HTTP server on random port.
- 04Feedback Confirmation and Taste Update
Summarizes preferred variant, ratings, and comments; confirms via AskUserQuestion before saving.