Build a Code‑Repository Q&A Agent Skill for OpenCode: From Installation to Custom Prompt Design

Overview

OpenCode is an open‑source AI coding assistant. By installing OpenCode, the DeepWiki MCP and the meta‑skill skill‑creator, you can build a custom Agent Skill that enables the model to answer technical questions about a set of open‑source repositories (vLLM, vLLM‑Ascend, MindIE‑LLM, MindIE‑SD, MindIE‑Motor, MindIE‑Turbo, msModelSlim).

Installation

Install OpenCode: curl -fsSL https://opencode.ai/install | bash Then add ~/.opencode/bin to PATH.

Install the meta‑skill skill‑creator (automatically via OpenCode or manually):

mkdir -p ~/.config/opencode/skills/skill-creator
cp -r /path/to/skill-creator/* ~/.config/opencode/skills/skill-creator/

Install DeepWiki MCP: opencode mcp add Verify with /mcps.

Creating the code-repos-expert Skill

In OpenCode’s Plan mode run:

/skill-creator 创建一个新的代码仓库智能问答 Skill —— code-repos-expert，并支持中文（Chinese）和英语（English）回答。Skill 内容得是英文。

Provide the following specification (English only):

1 - Ascend inference ecosystem code‑repository Q&A expert: answer usage, deployment, supported models, features, architecture, configuration, debugging, testing, performance optimization, custom development, source‑code analysis, etc. Support bilingual replies.
2 - For each repository mentioned, use DeepWiki to query owner/repo + intent‑derived query.
3 - If the repository cannot be determined, ask the user for clarification and never guess. Mark uncertain information with a disclaimer.

Target repositories:

https://github.com/vllm-project/vllm

https://github.com/vllm-project/vllm-ascend

https://gitcode.com/Ascend/MindIE-LLM

https://gitcode.com/Ascend/MindIE-SD

https://gitcode.com/Ascend/MindIE-Motor

https://gitcode.com/Ascend/MindIE-Turbo

https://gitcode.com/Ascend/msmodelslim

After the plan phase, switch to Build mode and confirm: 继续创建此 Skill The generated folder ~/.config/opencode/skills/code-repos-expert can be activated by copying it into the OpenCode skills directory.

Skill Design

1. Intent‑Mapping Table

Natural‑language expressions are mapped to technical intents (e.g., “太慢了” → Performance optimization, “报错了” → Troubleshooting). The table is stored in SKILL.md and drives intent identification.

2. Repository Routing Table

Keywords are mapped to the corresponding GitHub repository. Example: the keyword vllm‑ascend triggers a dual query – first the upstream vllm-project/vllm, then the plugin vllm-project/vllm-ascend. The full routing table is included in the skill.

3. Disambiguation Protocol

If the user mentions an ambiguous term (e.g., “MindIE” or “vllm” without “ascend”), the agent asks a clarifying question instead of guessing.

4. Hallucination Guard

When DeepWiki returns no result, the skill explicitly states that the information is unavailable and adds a disclaimer such as “(This information may be uncertain – please verify with official documentation).”

Typical Usage Examples

Example 1 – Deep Technical Query

/code-repos-expert vllm-ascend 具体是怎么结合 vllm 来适配 Qwen3-Next 的？必须深入分析关键的模型 patch 和算子适配 patch，并重点关注 patch_triton 中的具体内容

The agent first queries vllm-project/vllm for architecture details, then vllm-project/vllm-ascend for Ascend‑specific patches, and clearly marks the source of each answer.

Example 2 – Cross‑Repository Relationship

/code-repos-expert MindIE-LLM、MindIE-SD、MindIE-Motor、MindIE-Turbo 这四者之间的关系？

The skill aggregates information from all four repositories and returns a concise overview of their architectural connections.

Key Implementation Details

SKILL.md structure : YAML‑style metadata (name, description) followed by a Markdown body that contains the workflow, intent table, routing table, and response formatting rules.

DeepWiki commands used in the skill:

mcp__deepwiki__ask_question(repoName="owner/repo", question="...")

and optional structure queries mcp__deepwiki__read_wiki_structure or mcp__deepwiki__read_wiki_contents.

Response format : conclusion first, then detailed steps; all code snippets, file paths, and configuration names are presented verbatim; source attribution is included for each piece of information.

Limitations

The skill only covers the seven repositories listed above. Queries outside this scope are rejected with a clear statement that the skill does not support them.

Reference

Full source code and SKILL.md are available at https://github.com/Agent-Skill-007/learn-agent-skills.