temidayoxyz/deep-design

Deep Design is a self-hosted design-mode plugin purpose-built for DeepSeek Harness, positioning itself as a local alternative to Claude Design, Open Design, Google Stitch, and Figma Make, and it runs entirely on the DSH agent engine installed on the user's machine. The repository ships a single plugin that adds three coordinated pieces into the user's roster: a fully wired Deep Design agent preset (shell, files, web research, skills, plans, subagents, workflows), a deep-design-principles skill that encodes first-principles briefs, design-language-first thinking, anti-slop audits, and a scoring rubric, and a deep-design-qa skill that drives the visual feedback loop with headless renders at desktop and mobile breakpoints, a dedicated critic subagent, targeted fixes, and pixel-sampling fallback. Installation is delivered through the standard DSH surface command — dsh plugin --profile web add dsh-deep-design once it is published on the npm registry, or dsh plugin --profile web add ./deep-design for a checkout — and the preset lands in the canonical $DSH_HOME/.agent-presets/design/ directory every DSH deployment already scans. The install is declared idempotent and never overwrites an existing user-edited preset.

The product targets designers and indie developers who want a fully local, model-agnostic design studio instead of a hosted cloud product, and its loop mirrors the brief → direction → design system → artifact → render → critic → fix → re-render → compare → deliver pattern that cloud equivalents expose. Critically, the skill catalog entries for deep-design-principles and deep-design-qa register globally, so any DSH agent mode — not only Deep Design sessions — can pull them in when an upstream task genuinely needs design judgment or visual QA. Design memory compounds across sessions because the loop reads .design/ in the workspace at session start and writes lessons at session end, so output quality improves with continued local use rather than depending on a smarter hosted model.

Dependencies and limits are spelled out in the README: a vision-capable model is strongly recommended for Deep Design sessions because the critic benefits from actual image input, but the pixel-sampling fallback keeps the loop honest when only text-only models are available; web research is turned on at the preset level (fetch: true), so users must configure the DSH web provider before grounding references against outside material. The whole pipeline runs locally — nothing leaves the machine unless the user explicitly picks a cloud model — and both the preset and the skills live entirely under standard DSH directories so uninstalling is just a directory removal plus a Web UI restart. The project is released under the MIT license, the repository comes in at two GitHub stars and exposes a bundle-manifest declaration, and the first-run checklist is to reinstall the plugin with the surface used by the user's DSH profile, restart the Web UI, then pick the Deep Design chip on the new-session screen.