Version 1 of the prompt is simple
Instruction: Create a problem statement to solve a problem that we are currently facing because of context limitations in LLMs, but will go away at some point. The discovery of relevant tools in a composition of tools exposed via MCP servers either at the MCP framework level or in the agent itself.
Context:
Existing art : Lots of people making MCP servers, some local, some remote. On the MCP client, you have to selectively select and unselect tools.
Ideas: Allow one set of mcp tools to set the capabilities, needed, whereas another set is malleable through generally applicable sets of tool verbiage but in different context. e.g. CRUD for databases or files,
Ideas: After a MCP composition is discovered, allow user to toggle which tools are on or off, via a web interface, which stores the data in JSON (stored in a table for persistence), but can also be persisted , retrieved from a YAML file
Ideas: After a MCP composition is discovered, allow user to edit prompts, system instructions, via a web interface, which stores the data in MARKDOWN, JSON, YAML(stored in a table for persistence), but can also be persisted , retrieved from a MARKDOWN / folder structure
Format:Markdown
In deep research or extended thinking, the AI may ask back
**To create a well-targeted problem statement, could you clarify a few things:
Who is the primary user or target audience for this system (developers, agents, end-users)?
Should the focus be on tool discovery, tool selection UX, tool metadata structure, or agent composition?
Do you want the problem statement to reflect the current limitation specifically due to LLM context limits (e.g. prompt/token size, serialization issues), or more broadly the system complexity?
Is the goal to eventually eliminate this issue through better agent/MCP coordination, more expressive prompt representations, or persistent metadata APIs?
With that, I can frame the problem sharply and in context.**
Then I answered
**You said:
1. developers, agents, end-users
2. tool metadata structure, then tool selection UX, then tool discovery
3. current limitation due to constraints is one problem, the configuration of it in code / nocode , database, is really a matter of a circuitry making which is how we would say we are wiring robots, so a bit of it needs to be about a longer term strategy for design / management of these things
4. focus on the best way to allow people to compose their customized versions of MCP compositions**
Then it went to work. I then asked Claude Pro to improve the conversation and make me a better prompt.
# Improved Prompt for MCP Tool Composition Problem Statement
## Instruction
Create a comprehensive problem statement that addresses the challenge of managing and composing MCP (Model Context Protocol) server tools in a way that:
1. **Solves the immediate context limitation problem**: Address how current LLM context windows restrict the number of tools that can be exposed simultaneously, making tool selection and discovery inefficient
2. **Establishes a sustainable composition framework**: Design a system that treats MCP tool composition as "circuit design" or "robot wiring" - a configurable, persistent, and reusable approach to tool orchestration
3. **Serves multiple stakeholders**: Support developers (who create tools), agents (who consume tools), and end-users (who need intuitive interfaces) with appropriate abstractions for each
The problem statement should propose a solution architecture that prioritizes:
- **Tool metadata structure** (standardized, expressive, queryable)
- **Tool selection UX** (dynamic toggling, context-aware filtering)
- **Tool discovery** (semantic search, capability matching)
## Context
### Current State
- **Proliferation of MCP servers**: Developers are creating numerous MCP servers (local and remote) with varying capabilities
- **Manual tool management**: MCP clients require manual selection/deselection of tools, leading to cognitive overhead
- **Context window pressure**: LLMs have limited context windows, making it impractical to expose all available tools simultaneously
- **Lack of standardization**: No unified approach to tool metadata, discovery, or composition
### Emerging Patterns
- **Generic tool patterns**: Common operations (CRUD for databases/files) appear across different contexts with similar verbiage but different implementations
- **Capability-based selection**: Some tools define core capabilities while others are contextually adaptable
- **Need for persistence**: Tool configurations need to be saved, versioned, and shared across sessions and users
### Proposed Solution Elements
- **Web-based configuration interface**: Allow runtime toggling of tools with state persistence in JSON/database
- **Prompt and instruction editing**: Enable customization of system instructions and prompts through web UI
- **Multiple persistence formats**: Support JSON, YAML, and Markdown for different use cases (configuration, documentation, sharing)
- **Hierarchical tool organization**: Folder structures for logical grouping and discovery
## Format
Generate a **Markdown document** with the following structure:
```markdown
# Problem Statement: [Title]
## Executive Summary
[2-3 paragraphs summarizing the core problem and proposed solution]
## Problem Definition
### Current Limitations
- [Specific context limitation issues]
- [Tool management challenges]
- [Discovery and selection pain points]
### Long-term Vision
[Description of the "circuit design" approach to MCP composition]
## Proposed Solution Architecture
### 1. Tool Metadata Layer
[Standardized schema for tool description, capabilities, requirements]
### 2. Composition Engine
[How tools are selected, filtered, and presented to LLMs]
### 3. Configuration Management
[Persistence, versioning, and sharing mechanisms]
### 4. User Interfaces
#### For Developers
[Tool creation and metadata authoring]
#### For Agents
[Programmatic access and capability queries]
#### For End-users
[Visual composition and toggle interfaces]
## Implementation Roadmap
### Phase 1: Immediate Context Relief
[Solutions that work within current LLM limitations]
### Phase 2: Composition Framework
[Building the "wiring" system for tool orchestration]
### Phase 3: Ecosystem Integration
[Scaling to support community-driven tool libraries]
## Success Metrics
- [Measurable outcomes]
## Open Questions
- [Key decisions to be made]
Ensure the problem statement: