Agent Skills for DeepSeek V4: SKILL.md with Pro and Flash
Agent skills for DeepSeek V4: install SKILL.md in .agents/skills or .codex/skills, pair with DESIGN.md, and use Flash vs Pro without relying on screenshots.
Agent skills for DeepSeek V4 are standard SKILL.md folders your Codex-compatible or OpenAI-compatible harness already scans, usually under .agents/skills/ or .codex/skills/. V4 does not have a separate Moonshot-style skills product. You install the same design skills you use for Codex, then write the body for a text-only model: numeric checks, file scope, and "ask if not in DESIGN.md."
Skills are the procedure. Templates are the look. Hub: what are agent skills?.
Why V4 skills look different from Kimi or Cursor skills
Kimi K3 skills can demand a screenshot. Cursor + Grok skills can @ a PNG. V4 skills that say "look at the preview" will be skipped or hallucinated.
Rewrite the visual steps as measurable gates:
- "Confirm
h1uses the DESIGN.md display size, nottext-7xl." - "Do not add
bg-gradient-*classes." - "Stop after the named files. Print a diff summary."
That is still a skill. It is just a skill a blind agent can pass.
V4 also keeps reasoning across tool-call turns. A skill that says "read DESIGN.md first, then plan, then edit" actually compounds. Cheap Flash traces are long enough for that sequence to matter.
Where to put SKILL.md for DeepSeek V4
Depends on the harness, not on DeepSeek's weights.
| Harness | Project skills | User skills |
|---|---|---|
| Codex / ChatGPT coding | .agents/skills/, .codex/skills/ | ~/.agents/skills/ |
| OpenAI-compatible CLI pointed at DeepSeek | Same as that CLI's docs | Same |
| Cursor custom model → V4 | .cursor/skills/, often also .agents/skills/ | Cursor user skills dir |
July 31, 2026: DeepSeek-V4-Flash API is adapted for Codex. If you already install skills for Codex, Flash is the model you swap in. Paths do not change.
Prefer committing .agents/skills/ so Kimi Code and Cursor can share the same taste skill. See install skills across agents.
A V4-safe design skill
---
name: tasteful-ui
description: Use when building or redesigning UI, landing pages, dashboards, or marketing sections with DeepSeek V4. Enforces DESIGN.md, numeric spacing, section scope, and anti-slop bans. Do not use for backend-only work.
---
# Tasteful UI (DeepSeek V4)
1. Read DESIGN.md. If absent, stop.
2. Restate tokens and anti-patterns before editing.
3. Implement one named section and named files.
4. Never use screenshots as a requirement. Use type steps, gaps, and class names.
5. If a visual choice is not in DESIGN.md, ask. Do not invent.
6. Prefer existing files under components/ui/. No new font imports.
7. End with a file list and what changed. Do not start the next section.
Invoke explicitly in Codex-style clients when auto-trigger is flaky:
Use skill tasteful-ui.
Model: deepseek-v4-flash.
Hero only. Files: components/landing/hero.tsx, app/page.tsx.
Which design skills to install for V4
From /skills, same priority as other models, with one filter: skip skills whose only gate is "compare to the screenshot."
| Skill type | V4 note |
|---|---|
| Taste / anti-slop | Highest leverage; bans are text (taste skills) |
| Landing / pricing | Section recipes Flash can run as a checklist |
| Typography | Replace vision QA |
| Color system | Role names V4 can grep for in class lists |
| Motion | Duration caps in writing |
| Accessibility | Contrast numbers, focus rings; V4 will not "see" a fail |
Category overview: best design agent skills 2026.
Flash vs Pro when a skill is on
| Situation | Model |
|---|---|
| Skill-following section work | V4 Flash |
| Skill + huge repo + many tools | V4 Pro |
| Skill ignored, agent inventing UI | Re-read skill, then Pro, still one section |
| Codex day-to-day | Flash (Codex-adapted) |
Think Max is a reasoning mode, not a skill. Do not turn Max on to compensate for a missing DESIGN.md.
Tool schema vs skills
V4's |DSML| XML tools are a harness problem. Skills should not re-teach the XML format unless you maintain a custom runner. Keep SKILL.md about craft. Keep DSML in the client.
If tool calls fail, the skill never runs. Fix the harness first.
AGENTS.md alongside V4 skills
Always-on file for the repo:
# AGENTS.md
- DESIGN.md is the only source of visual tokens.
- One section per request.
- No gradients, no new fonts, no extra routes.
- Ask when the spec is silent.
Skills trigger on UI tasks. AGENTS.md catches the backend-adjacent "while I'm here I'll restyle the nav" impulse, which Flash will happily do because it is cheap.
FAQ
Does DeepSeek V4 have native Agent Skills?
The model does not invent a new skill format. Your coding agent does. Point that agent at V4 and put SKILL.md where the agent already looks.
Can I use Kimi or Cursor skills as-is?
Yes if they are portable SKILL.md. Delete or rewrite screenshot-only steps. Keep DESIGN.md as the first read.
Why does V4 ignore the skill on Flash?
Description too vague, or the first user message never mentioned UI. Put "landing page" / "DESIGN.md" in the description. Mention the skill name in the prompt for the first few sessions.
Do I need a different skill for Pro and Flash?
No. One skill. Different model ids. Pro is for when Flash drops the procedure on a long trace.
Procedure for a model that cannot see
DeepSeek V4 will follow a skill that is written like a linter. Install one taste skill under .agents/skills/, commit DESIGN.md, and run Flash until the section exists.
Browse portable design skills on /skills and pair them with templates from the gallery.
Ship the next screen with taste
Browse agent-ready templates, DESIGN.md specs, and prompts in the gallery — then paste into Cursor, Claude Code, or v0.
