# Case Study: Recursive Prompt Engineering
> Status: narrative
Built a prompt engineering tool. Used it to build itself.
---
## The Problem
Prompt iteration is slow. Edit → restart → test → repeat. Each cycle: 5-10 minutes.
- Prompts scattered across projects, no version control
- Output quality varies without structured reasoning
- Quality validation is manual
## The Solution
**Claude Prompts MCP Server**—a hot-reloadable prompt engine with structured execution.
```bash
npx -y claude-prompts@latest
```
```
prompt_manager(action:"list") # See your prompts
prompt_engine(command:">>analyze_code @CAGEERF") # Run with methodology
```
| Before | After |
| -------------------------------- | ------------------------------------------ |
| Edit → restart → test (5-10 min) | Edit → test instantly (<1 min) |
| Scattered prompt files | Git-versioned Markdown library |
| Variable reasoning quality | 4 methodology frameworks enforce structure |
| Manual quality checks | Quality gates auto-validate |
---
## Demo
```bash
# List prompts
prompt_manager(action:"list")
# Run with methodology
prompt_engine(command:">>diagnose @CAGEERF scope:'performance'")
# Chain steps
prompt_engine(command:"Analyze code --> Identify issues --> Propose fixes")
# Inline quality gates
prompt_engine(command:"Summarize :: 'under 200 words' :: 'include stats'")
```
---
## Technical Architecture
**Pipeline-First Design**: Every request flows through a unified 21-stage pipeline.
```
Parse → Validate → Plan → Enhance → Execute → Format → Return
```
- **Predictable**: Same path for every request. When something fails, you know exactly where.
- **Extensible**: New feature = new stage. No rewiring.
- **Debuggable**: LLM interactions need deterministic flow.
**Symbolic Command Language**:
```
>>analyze_code @CAGEERF :: 'security review' --> >>summarize
Operators:
>> Execute prompt by ID
@ Apply methodology (CAGEERF, ReACT, 5W1H, SCAMPER)
:: Quality gate (inline criteria)
--> Chain to next step
```
**Hot-Reload**: Edit a prompt file, changes apply immediately. No restart.
---
## Results
| Metric | Value |
| -------------- | ----------------------- |
| GitHub Stars | 115+ |
| Forks | 27 |
| Iteration Time | 5-10 min → <1 min (10x) |
---
## The Recursive Part
This case study was refined using the tool itself:
```
prompt_engine(command:">>readme_improver context:'case-study.md' :: 'must include metrics'")
```
When a prompt underperforms:
```
User: "The output is too verbose"
Claude: prompt_manager(action:"update", id:"readme_improver", ...)
User: "Test it"
Claude: prompt_engine(command:">>readme_improver") # Updated version, instantly
```
No manual editing. No restart. The tool improves its own prompts.
**Meta-Prompts: Prompts That Create Prompts**
```
>>create_gate name:"API Docs"
```
Two-phase UX:
1. **Design phase** — Missing fields? Template shows guidance and examples
2. **Validation phase** — All fields present? Script validates → auto-creates
Same pattern for `>>create_prompt` and `>>create_methodology`. Zero API memorization.
---
## Key Decisions
| Decision | Trade-off |
| ------------------------------------------- | ------------------------------------------------------------ |
| **Unified pipeline** over per-tool handlers | Higher complexity upfront, but predictable debugging |
| **File-based state** over database | No dependencies, survives restarts, but single-instance only |
| **Symbolic DSL** over JSON-only API | Learning curve, but natural prompt composition |
---
## Architectural Evolution
- **Scattered → Unified**: First iteration had individual handlers. Debugging was unpredictable. Single pipeline made LLM interactions deterministic.
- **Grow → Understand → Decompose**: Components started broad, grew while discovering requirements. Premature abstraction creates wrong boundaries—decomposition happened once the domain was clear.
- **DSL emerged from usage**: Started JSON-heavy. After watching real composition patterns, operators became obvious. Extracted, not designed upfront.
- **Gates as core**: Quality validation was initially bolted on. Promoting gates to first-class (own registry, hot-reload, activation rules) made the difference between "prompts that might work" and "prompts that reliably work."
---
## Links
- [GitHub](https://github.com/minipuft/claude-prompts-mcp)
- [NPM](https://www.npmjs.com/package/claude-prompts)
- [Architecture](../architecture/overview.md)
- [Design Decisions](design-decisions.md)
