スキル一覧に戻る
wildwasser

agent-tuning

by wildwasser

Multi-agent AI development templates for opencode

1🍴 0📅 2026年1月20日
GitHubで見るManusで実行

SKILL.md


name: agent-tuning description: Configure and optimize AI coding agents (OpenCode/Claude). Use when setting up or improving agent behavior.

Configuration Locations

OpenCode

~/.config/opencode/
├── opencode.json          # Main config
├── agent/                 # Custom agent prompts
│   ├── oscar.md
│   ├── scout.md
│   ├── ivan.md
│   └── jester.md
├── skills/                # Reusable skills
│   └── <name>/SKILL.md
└── rules/                 # Always-active rules
    └── RULES.md

Claude Code

~/.claude/
├── settings.json          # Main config
├── agents/                # Custom agents
│   └── <name>.md
├── skills/                # Reusable skills
│   └── <name>/SKILL.md
└── CLAUDE.md              # Global instructions

Agent File Structure

Frontmatter (YAML)

---
tools:
  read: true
  write: false
  edit: false
  glob: true
  grep: true
  bash: true
  task: false
  webfetch: true
  todoread: true
  todowrite: true
  question: true    # Ask user questions
  skill: true       # Load skills

permission:
  bash:
    "git *": allow
    "gh *": allow
    "*": deny
---

Tool Reference

ToolPurposeRisk Level
readRead filesLow
writeCreate/overwrite filesHigh
editModify filesHigh
globFind files by patternLow
grepSearch file contentsLow
listList directoriesLow
bashExecute commandsVariable
taskDelegate to subagentsLow
webfetchFetch web contentMedium
todoreadRead task listLow
todowriteUpdate task listLow
questionAsk user questionsLow
skillLoad skill filesLow

Permission Patterns

Restrictive (Read-Only Agent)

tools:
  read: true
  write: false
  edit: false
  bash: true

permission:
  bash:
    "ls *": allow
    "cat *": allow
    "git status": allow
    "*": deny

Permissive (Full Access)

tools:
  read: true
  write: true
  edit: true
  bash: true

permission:
  bash:
    "*": allow

Balanced (Common Pattern)

tools:
  read: true
  write: true
  edit: true
  bash: true

permission:
  bash:
    # Safe commands
    "ls *": allow
    "cat *": allow
    "head *": allow
    "tail *": allow
    
    # Git operations
    "git *": allow
    
    # Python in venv only
    ".venv/bin/*": allow
    
    # Deny everything else
    "*": deny

Agent Roles

Orchestrator (Oscar-style)

  • Has task: true to delegate
  • Has question: true to interact with user
  • Limited direct tool access
  • Coordinates other agents

Researcher (Scout-style)

  • Read-only access
  • Has webfetch: true
  • Full search tools (glob, grep)
  • No task (can't delegate)

Implementor (Ivan-style)

  • Full write access
  • Full bash access
  • No task (can't delegate)
  • Does the actual work

Reviewer (Jester-style)

  • Read-only access
  • Different model for diversity
  • Challenges assumptions
  • No write access

Tuning Tips

1. Start Restrictive

Begin with minimal permissions, expand as needed.

2. Use Specific Bash Patterns

# Good - specific
".venv/bin/pytest *": allow

# Bad - too broad
"*pytest*": allow

3. Separate Concerns

  • Orchestrator: coordinates
  • Researcher: investigates
  • Implementor: changes code
  • Reviewer: validates

4. Test Permission Changes

After modifying permissions, test that:

  • Allowed commands work
  • Denied commands are blocked
  • Edge cases are handled

Debugging

Check Active Config

# OpenCode
cat ~/.config/opencode/opencode.json

# View agent prompt
cat ~/.config/opencode/agent/<name>.md

Common Issues

  • Tool not working: Check tools: section
  • Command denied: Check permission.bash patterns
  • Skill not loading: Check skill name matches directory
  • Agent not found: Check opencode.json agent definitions

スコア

総合スコア

50/100

リポジトリの品質指標に基づく評価

SKILL.md

SKILL.mdファイルが含まれている

+20
LICENSE

ライセンスが設定されている

0/10
説明文

100文字以上の説明がある

0/10
人気

GitHub Stars 100以上

0/15
最近の活動

3ヶ月以内に更新がある

0/10
フォーク

10回以上フォークされている

0/5
Issue管理

オープンIssueが50未満

+5
言語

プログラミング言語が設定されている

+5
タグ

1つ以上のタグが設定されている

0/5

レビュー

💬

レビュー機能は近日公開予定です