DeepSeek Deep Code skills guide: use the official skill folders and keep project-specific instructions local

DeepSeek's Deep Code page says Agent Skills are discovered from folders containing `SKILL.md` files. This is not a vague 'plugin' story. It is a filesystem contract the tool already understands.

That matters because it tells developers to think in reusable instruction packs rather than ad hoc prompt snippets copied into every session.

DeepSeek documents two discovery roots for Agent Skills: user-level skills under `~/.agents/skills/<name>/SKILL.md` and project-level skills under `./.deepcode/skills/<name>/SKILL.md`.

That split is useful because it gives you one place for personal reusable workflows and another place for repository-specific instructions that should travel with the codebase.

A practical rule follows from the official folder split: if a skill assumes one repository layout, one deployment flow, or one coding standard, keep it in `./.deepcode/skills` so the instructions stay versioned with that project.

Use `~/.agents/skills` only for genuinely reusable personal workflows that make sense across many repositories. Otherwise you turn a repo-specific shortcut into global hidden state.

This matters for search intent because many users are really asking 'where should this instruction live so my teammates see it too'. The project-local Deep Code path is the cleaner answer.

DeepSeek's page says you can press `/` to open the skills and commands menu, or type the skill name directly. That means the supported discovery surface is already in the TUI.

When a new skill appears not to work, the right first check is whether it shows up in the slash-command picker. If it does not, the issue is usually folder placement or file naming rather than model behavior.

The official docs do not prescribe a giant skill taxonomy, and that restraint is useful. Start with one narrow skill that solves a recurring workflow, such as local build verification, release-note drafting, or repository-specific lint triage.

Once one skill is clearly useful, split later skills by job rather than building a monolithic instruction file that tries to cover everything. Smaller skills are easier to test and less likely to inject the wrong policy into unrelated work.

This is a support page for a DeepSeek-native coding workflow, not a new product listing. It should not create plan cards or imply that every tool in the agent ecosystem is sold here.

For the base CLI wiring, compare `/guides/deepseek-deep-code-cli`. For other official DeepSeek agent surfaces, compare `/guides/deepseek-opencode-setup`, `/guides/deepseek-github-copilot-cli`, and `/guides/deepseek-openclaw-wechat-feishu`.