discuss-mcp

Discuss MCP

Advisory skill for reflecting on and improving an MCP server. Unlike builder skills (/build-tools, /build-evals) that execute generation workflows, this skill is conversational — it helps developers understand what to improve, why, and how.

When to Use

Use this skill when the developer asks:

• "How can I improve my MCP server?"
• "Why are my evals slow/expensive/failing?"
• "Should I add composite tools or proxy tools?"
• "What tools am I missing?"
• "What should I build next?"
• "Are my tools well-designed?"
• "How do I reduce eval cost/turns?"

Improvement Paths

Route to the appropriate sub-file based on what the developer wants to discuss:

1. Trace Optimization — "My evals are slow/expensive/failing"

The developer wants to analyze eval performance and iteratively improve tools based on trace data.

Read: trace-optimization.md

Signals: mentions evals, traces, cost, turns, tokens, performance, "why does the LLM struggle", repeated tool calls, wrong tool selection.

2. Chained Tool Design — "Should I add composite/proxy tools?"

The developer wants to discuss adding tools that delegate to other MCP servers — whether to proxy or create composites.

Read: chained-tools.md

Signals: mentions chaining, proxy, composite, multi-server, delegation, orchestration, combining data from multiple sources.

3. Behavioral Analysis — "What should I build next?"

The developer wants a holistic review of their MCP: coverage gaps, organizational quality, prioritization of next steps.

Read: behavioral-analysis.md

Signals: mentions coverage, gaps, what's missing, next steps, priorities, tool organization, modes, slices, "are my tools well-designed".

Conversational Guidelines

When advising on MCP improvements, follow these principles:

1. Ask before recommending. Understand the developer's goals before suggesting changes. "What workflows are you trying to support?" is more useful than jumping to solutions.

2. Propose with trade-offs, not prescriptions. Present options with pros and cons. "You could simplify the response schema (reduces tokens but loses detail) or add a composite tool (preserves detail but adds complexity)."

3. Use their own codebase as evidence. Reference specific tool files, descriptions, eval results, and collection structures. Abstract advice is less useful than "your get-document-by-id description says X but the eval trace shows the LLM tried Y first."

4. Enter plan mode for concrete changes. When a specific improvement is identified and the developer agrees, enter plan mode to design the change before executing. This ensures the developer approves the approach.

5. Summarize action items. After each discussion thread, list what was agreed: changes to make, things to investigate, next steps. Keep it concrete.

6. Iterate, don't overhaul. Prefer small, measurable improvements over large rewrites. One description fix that drops turns from 5 to 3 is more valuable than a speculative restructuring.

7. Respect ignored endpoints. Read docs/analysis/IGNORED_ENDPOINTS.md before suggesting new tools. Endpoints listed there are deliberately excluded — do not recommend implementing them. They represent settled decisions about scope.

Multiple Topics

If the developer's question spans multiple paths (e.g., "my evals are slow and I think I need composite tools"), read both sub-files and synthesize. Start with the most pressing concern — usually trace optimization reveals whether chaining is actually needed.