deep-design

Deep Design

Design mode for DeepSeek Harness — a self-hosted alternative to Claude Design, Open Design, Google Stitch, and Figma Make, built on the DeepSeek Harness agent engine.

Deep Design turns your local harness into a design studio. It runs the same loop those products run — brief → direction → artifact → critique → deliver — with rendered visual QA at every pass, on your machine, with your own models and API keys.

What it installs

One plugin, three things:

1. Deep Design — an agent preset in your roster: a complete design agent (shell, files, web research, skills, plans, subagents, workflows) whose entire job is the design loop.
2. deep-design-principles — the design intelligence: first-principles brief, design language before implementation, composition, motion, anti-slop audits, a scoring rubric, and a final self-review.
3. deep-design-qa — the visual feedback loop: headless render at desktop and mobile → a critic pass (a subagent that diagnoses but never writes code) → targeted fixes → re-render → compare. With a pixel-sampling fallback for models without image input.

Both skills register globally, so any agent mode can load them from the skill catalog — not just Deep Design sessions.

Install

dsh plugin --profile web add dsh-deep-design    # from the npm registry once published, or
dsh plugin --profile web add ./deep-design      # from a checkout

Restart the Web UI, open the new-session screen, and choose Deep Design from the agent-preset chip. (The web profile is the one dsh web boots; use the same command with another profile name for other surfaces.)

The plugin installs the preset into $DSH_HOME/.agent-presets/design/, the standard user-preset directory every deployment scans. Installation is idempotent and never overwrites an existing directory — once you edit your copy, it is yours.

The loop

brief → direction (.design/direction.md) → design system → artifact
     → render (headless, desktop + mobile)
     → critic pass (structured visual diagnosis)
     → fix → re-render → compare → deliver

Design memory compounds across sessions: the loop reads .design/ in the workspace at session start and records lessons at session end. Designs get better with use, not just with better models.

Notes

• Best with a vision-capable model for Deep Design sessions — the critic sees better with eyes. The pixel-sampling fallback keeps the loop honest without one.
• Web research is enabled on the preset (fetch: true); configure the harness's web provider for reference grounding.
• Deep Design runs entirely on your machine. Nothing leaves it unless you choose a cloud model.

License

MIT.