Agent Skills for Kimi K3: SKILL.md in Kimi Code (2026)

Agent skills for Kimi K3 are SKILL.md packages Kimi Code discovers at project or user scope, then loads when the task matches the skill description. Put design skills in .kimi-code/skills/ or .agents/skills/, invoke with /skill:name when you want the procedure forced, and pair every taste skill with a DESIGN.md from the gallery so K3 does not improvise tokens.

Skills tell K3 how to work. Templates tell it what good looks like. You want both. Hub: what are agent skills?.

Why Kimi K3 needs skills more than a smaller model

K3 is trained for long, hard tasks. Moonshot warns it can be over-proactive: if the brief is thin, it decides for you. A skill with explicit stop conditions is how you keep a 2.8T model inside the section you named.

Skills also survive model switches. The same SKILL.md can live in .agents/skills/ for Kimi Code, Cursor, and Codex. That matters if you draft in K3 and finish in Grok 4.6.

Where Kimi Code looks for skills

Kimi Code scans scopes in order: Project > User > Extra > Built-in. More specific wins on name conflict.

Scope	Paths (typical)
Project	`.kimi-code/skills/`, `.agents/skills/` (from the git root)
User	`$KIMI_CODE_HOME/skills/` (default `~/.kimi-code/skills/`), `~/.agents/skills/`
Extra	`extra_skill_dirs` in Kimi Code `config.toml`

Directory form is the one to use:

.kimi-code/skills/tasteful-ui/SKILL.md
.kimi-code/skills/tasteful-ui/references/

Flat tasteful-ui.md also works. The folder form wins if both exist. Official docs: Kimi Code skills.

Team pattern: commit .agents/skills/tasteful-ui/ and symlink or copy into .kimi-code/skills/ if you want the Kimi-specific path populated too. Cursor and Codex already read .agents/skills/.

SKILL.md that actually steers Kimi K3

Frontmatter is the retrieval key. Write the description like a query K3 would match.

---
name: tasteful-ui
description: Use when building or redesigning UI, landing pages, dashboards, or marketing sections. Enforces DESIGN.md tokens, section scoping, screenshot checks, and anti-slop bans. Do not use for backend-only tasks.
---

# Tasteful UI (Kimi K3)

1. Read DESIGN.md at repo root. If missing, stop and ask.
2. Confirm anti-patterns before any edit.
3. Implement one named section and named files only.
4. Do not invent fonts, colors, routes, or dependencies.
5. If intent is ambiguous, ask. Do not guess.
6. After UI changes, request a screenshot of the preview before the next section.

The "ask, don't guess" line is not decoration. It is the K3-specific control for over-proactivity.

Invoke when auto-trigger is not enough:

/skill:tasteful-ui
Implement the hero only per DESIGN.md and PROMPT.md.

Design skills to install first for Kimi K3

Install from /skills. Priority order for this model:

Skill type	Why K3 needs it
Taste / anti-slop	Attacks the median SaaS prior; see taste skills
Landing / pricing	Named section recipes so a long session does not invent a fourth feature row
Typography	Stops Inter-by-default even when vision says "looks fine"
Motion	Caps animation; K3 will otherwise "improve" hover states
Accessibility	Contrast and focus after the screenshot loop

Full category list: best design agent skills 2026.

Do not install twelve skills on day one. One taste skill plus DESIGN.md beats a crowded skills prompt.

Install steps (Kimi Code + Kimi K3)

Create .kimi-code/skills/tasteful-ui/ (or .agents/skills/tasteful-ui/).
Add SKILL.md with name + description + the steps above.
Copy a DESIGN.md from the gallery to repo root.
Start a new Kimi Code session. /model → Kimi K3.
Ask "What skills do you have?" or run /skill:tasteful-ui.
Run a small UI task. Confirm the plan cites DESIGN.md and stop conditions.

If the skill does not appear, you are not at the git root, or a user-level skill with the same name is winning. Check scope order.

More install patterns across agents: install skills in Claude, ChatGPT, Cursor, Kimi.

Skills + vision: the K3-only loop

A skill can require a screenshot. Most models will skip that line. K3 can obey it.

Add to the skill body:

After each section, ask the user for a preview screenshot.
Do not start the next section until visual issues from the screenshot are fixed or explicitly skipped.

That turns the skill into a gate, not a suggestion. Pair with UI templates for Kimi K3 so there is a real page to photograph.

AGENTS.md vs skills vs DESIGN.md

File	Always on?	Job for K3
AGENTS.md	Yes	"Do not improvise. Ask."
SKILL.md	When triggered	UI procedure + screenshot gate
DESIGN.md	When the skill says to read it	Tokens and layout

K3's thinking-history limit still applies. Do not switch to K3 in the middle of a Claude or Grok session and expect skills to save you. New session, then skills.

FAQ

Does Kimi K3 support Agent Skills?

Yes, via Kimi Code. Project skills live under .kimi-code/skills/ or .agents/skills/. User skills under ~/.kimi-code/skills/ or ~/.agents/skills/. Invoke with /skill:name.

Can I reuse Claude skills with Kimi K3?

If they are standard SKILL.md folders, yes. Put them on a path Kimi Code scans. Tighten the body for K3: stop conditions and screenshot gates. Claude-oriented skills that assume CLAUDE.md only should also mention DESIGN.md.

Why does K3 ignore my skill?

Description does not match the task, or a higher-priority skill with the same name exists. Make the description include "landing pages," "UI," "DESIGN.md." Restart the session after adding files.

Do skills replace templates?

No. Skills are procedures. Templates are composition. K3 with a skill and no template still invents a layout. K3 with a template and no skill still over-builds.

Load a skill, then load a spec

Kimi K3 will work for hours. A skill is how you decide what those hours are for. Install one taste skill, commit DESIGN.md, start a clean Kimi Code session on Kimi K3.

Browse design skills with before/after demos on /skills, and pair them with templates from the gallery.