
tool-minimization
by abdelrahman146
Kyora is a SAAS service that helps social media store owners manage their business and finances
SKILL.md
name: tool-minimization description: 'MCP/tool selection rules and lane-default toolchains for Kyora Agent OS. Use when selecting tools for a task, configuring agent tool lists, or understanding when to use MCP servers. Triggers: tools, MCP, tool selection, least-privilege, Context7, Playwright, GitHub MCP.'
Tool Minimization Workflow
Operational guide for selecting and minimizing tools in Kyora Agent OS. Implements least-privilege tool selection and safe MCP usage.
When to Use This Skill
- Selecting tools for a new task
- Configuring agent tool lists
- Deciding when to use MCP servers
- Reducing tool overload in agents/prompts
- Understanding lane-default toolchains
Prerequisites
- Understanding of current task scope and risk level
- Knowledge of available MCP servers (if considering MCP)
- Access to agent/prompt files for tool list auditing
Core Principle: Least Privilege
Start with the minimum tools needed. Add more only when clearly necessary.
Tool overload hurts reliability:
- More tools = more confusion for the model
- More tools = higher chance of incorrect tool selection
- More tools = longer context
Tool Selection Protocol
Step 1: Start with Workspace Read/Search
Default starting tools:
read— Read file contentssearch— Find files and text (grep, glob)semantic_search— Fuzzy concept matching
These solve most discovery and planning tasks.
Step 2: Add Only What's Needed
| Need | Tool to Add |
|---|---|
| Modify files | edit |
| Run commands | execute |
| Run tests | tests / runTests |
| Track tasks | todo |
| Invoke agents | agent |
Step 3: Consider MCP Only When Beneficial
MCP tools add latency and complexity. Use only when they clearly reduce work or increase certainty.
Lane-Default Toolchains
| Lane | Default Tools | Add Only If Needed |
|---|---|---|
| Discovery | read, search | GitHub MCP (repo context), Context7 (library uncertainty) |
| Planning | read, search | Context7 (API usage), diagrams only if asked |
| Implementation | read, search, edit, execute, tests | Playwright (UI), Context7 (new dep) |
| Validation | execute, tests | Playwright (UI smoke), DevTools (perf) |
| Recovery | read, search, get_changed_files | None unless blocked |
Role-Default Tool Lists
Orchestrator (Least Privilege)
tools: ['read', 'search', 'todo']
# edit only for plans/notes, NEVER production code
Leads
tools: ['read', 'search', 'edit'] # edit for specs only
# execute optional for validation
Implementers
tools: ['read', 'search', 'edit', 'execute']
# tests tool added as needed
QA/Test Specialist
tools: ['read', 'search', 'edit', 'execute']
# Playwright preferred for UI flows
Security Reviewer (Read-Only)
tools: ['read', 'search']
# NEVER edit or execute
MCP Server Policy
Approved MCP Servers
| Server | Purpose | When to Use |
|---|---|---|
| Context7 | Up-to-date library docs/snippets | New dependency; unfamiliar API |
| Playwright | UI automation, screenshots | UI flows, RTL verification, E2E |
| Chrome DevTools | Layout/perf/network debug | Performance issues, layout bugs |
| GitHub MCP | Issues/PRs/policies | Repo context (NOT local-only tasks) |
When NOT to Use MCP
- Simple local refactors
- Tasks solvable via workspace search + reading existing code
- Anything requiring secrets or sensitive logs
- Local-only tasks (don't use GitHub MCP)
MCP Safety Rules
See MCP Safety Checklist for detailed rules.
Key points:
- Treat local MCP servers as code execution
- Never paste secrets; use VS Code input variables
- Keep enabled tools under model/tool limits
Tool Correctness Rules
- Prefer workspace tools for local codebase truth
- Use Context7 only to confirm third-party API usage; stop after minimum needed
- Use Playwright/DevTools when visual/RTL/layout correctness matters
- Disable entire servers if not relevant to current task
Operational Notes
Tool List Caching
- Tool lists are cached by VS Code
- Reset cached tools when server tools change
- Use "Copilot: Reset Chat" or restart session
Tool Sets
- Group related tools into tool sets
- Keep enabled tools under model limits
- Disable unused servers to reduce picker noise
Success Checklist
- Started with read/search only
- Added tools only when clearly needed
- MCP tools have clear justification
- No unnecessary tools in final list
- Tool list matches role defaults
- Safety rules followed for MCP
- No surprise docs generated
- Stop-and-ask triggers checked
Stop-and-Ask Triggers
MUST ask before proceeding if:
- Adding MCP server not in approved list
- Tool requires secrets or sensitive data access
- Tool list exceeds role defaults significantly
- Uncertain if tool is necessary for task
Troubleshooting
| Issue | Cause | Solution |
|---|---|---|
| Tool not available | Not in enabled list | Check agent frontmatter tools: list |
| MCP server slow | Network latency or server load | Consider if workspace tools suffice |
| Tool list too long | Scope creep | Reset to lane defaults, add only essentials |
| Cached tools stale | Server tools changed | Use "Copilot: Reset Chat" or restart |
Validation Commands
No specific commands for tool selection, but verify:
# Check agent tool lists
grep -r "tools:" .github/agents/
# Check prompt tool lists
grep -r "tools:" .github/prompts/
References
SSOT References
- MCP + Tooling Policy: KYORA_AGENT_OS.md#L791-L879
- Tool selection in routing: KYORA_AGENT_OS.md#L477-L489
- Approved MCP servers: KYORA_AGENT_OS.md#L825-L845
Score
Total Score
Based on repository quality metrics
SKILL.mdファイルが含まれている
ライセンスが設定されている
100文字以上の説明がある
GitHub Stars 100以上
3ヶ月以内に更新がある
10回以上フォークされている
オープンIssueが50未満
プログラミング言語が設定されている
1つ以上のタグが設定されている
Reviews
Reviews coming soon