🇷🇺 Русская версия · English
Skills that turn an AI agent into a specialist for a concrete job — not a chat partner.
Each skill is a pack of instructions, templates, scripts, and reference material that drops into opencode (or any compatible CLI) and tells the agent exactly how to do one thing well: bootstrap a project, run a planned sprint, coordinate multiple models, or break out of a mode-collapsed loop.
No runtime. No lock-in. Just files the agent reads on demand.
| Skill | One-liner | Use when |
|---|---|---|
| project-bootstrap | Generates the agent "home" for a project in one session — AGENTS.md, handoff, memory, rules, adapted to project type and model. |
You're starting a new project or rescuing an existing one and want the agent infrastructure right. |
| wave-spec | A plan-gate skill for sprints and waves: INTENT → interview → SPEC/PLAN → approve → dispatch → lifecycle gates. Portable across OpenCode, ZCode, Qwen Code. |
You're running a multi-session sprint, a content/translation wave, or any work that must not skip planning. |
| multi-model-orchestration | Coordinates 2+ AI models (DeepSeek V4 Flash, Qwen 3.8 Max, GLM 5.2, GPT-5.5) for parallel review, cross-validation, or bulk work via Orca. | You need independent perspectives, a fidelity port, or a security/RLS review that one model alone can't gate. |
| vs-architect | Verbalized Sampling (arXiv 2510.01171) — generates diverse solution variants with probability estimates. | You're choosing between approaches, debugging an unknown root cause, or breaking out of a mode-collapsed loop. |
New project, "set up the agent structure" → project-bootstrap
Sprint/wave that must not skip planning → wave-spec
2+ models for review, cross-validation, bulk work → multi-model-orchestration
Diverse variants with probabilities → vs-architect
The three planning skills chain: project-bootstrap sets up the agent home → wave-spec runs the sprint inside it → multi-model-orchestration cross-reviews the work. vs-architect is a standalone thinking tool.
git clone git@github.com:dimkurilo/opencode-skills.git ~/Projects/opencode-skills
# symlink the skills you want into opencode
for skill in project-bootstrap wave-spec multi-model-orchestration vs-architect; do
ln -sfn ~/Projects/opencode-skills/skills/$skill ~/.config/opencode/skills/$skill
doneManual install (no symlinks): cp -R skills/<name> ~/.config/opencode/skills/<name>. opencode picks up new skills on the next launch.
Several skills also install into Grok (~/.grok/skills/) and other CLIs that follow the same SKILL.md convention.
skills/<name>/
├── SKILL.md # main instructions + YAML frontmatter (name, description)
├── README.md / README.ru.md
├── references/ # reference material, examples, theory
├── assets/templates/ # generation templates with ${VARIABLE} placeholders
└── scripts/ # helper shell/Python scripts (lint, verify, classify)
The frontmatter description tells the host agent when to load the skill. The body is the workflow. References are loaded on demand. Templates produce the artifacts (SPEC.xml, PLAN.xml, AGENTS.md, briefs, handoffs) the skill generates.
Same convention: a directory under skills/, a SKILL.md with frontmatter (name, description — describe when to use, not only what), optional references/, assets/templates/, scripts/. The skill-creator and skill-audit skills (in ~/.config/opencode/skills/) help you write and review them.
opencode-skills/
├── README.md / README.ru.md
├── CHANGELOG.md
├── LICENSE # MIT
└── skills/
├── project-bootstrap/
├── wave-spec/
├── multi-model-orchestration/
└── vs-architect/
Public, MIT-licensed. Inspired by PromptPasture/agent.md, Cursor Rules, OpenCode Rules, vv-opencode, and the Agent1st Protocol.
MIT