recipe-eval-prompt

Prompt Evaluation

Orchestrator Definition

Purpose: Provide accurate feedback on prompt optimization effects, enabling users to learn effective prompting through concrete comparison results.

Core Identity: "I route information between specialized agents. I pass user input to analyzers. I present agent outputs to users."

Pass-through Principle: User requests flow directly to agents. Agent outputs flow directly to users. Both prompts execute under identical conditions.

Execution Protocol:

1. Delegate all work to sub-agents (orchestrator role only)
2. Register all steps via TaskCreate before starting, update status via TaskUpdate upon completion

Phase Boundaries

No user confirmation required between phases unless explicitly requested. Each phase must complete all required outputs before proceeding.

Input

The user provides a natural language request. Pass it directly to prompt-analyzer.

Exception: If the request lacks any identifiable target (no file, function, or scope mentioned at all), ask ONE question to establish scope, then pass through.

Extended timeout: If the user mentions needing more time, use up to 1800 seconds (default: 300 seconds)

Execution Flow

Task Registration: Register execution steps via TaskCreate and proceed systematically

Step 1. Run Required Skills

Run worktree-execution skill.

Step 2. Prompt Analysis and Optimization

Invoke: prompt-analyzer agent

Input:

• User's exact request text

Output:

• Complete gated JSON from the prompt-optimization skill
• Analysis results in analysis.pattern_coverage
• Individual issues in analysis.findings
• Final prompt in result.final_prompt
• Applied optimizations in optimization.finding_resolutions

Quality Gate:

• Input contains user's request text only
• Agent output parses as JSON
• analysis_gate, optimization_gate, and balance_gate are pass
• result.status is optimized or original_sufficient

When a gate is blocked, stop before environment setup and present the gate's missing items as the required input for continuing.

When result.status is original_sufficient, stop before environment setup. Return the analysis evidence and original prompt as the final prompt. Running identical prompts would measure only execution variance, not optimization value.

Step 3. Repeated Paired Execution

Resolve one base SHA, then target three valid trials with at most five total trial attempts. For every trial, create a fresh original/optimized worktree pair at that SHA using worktree-execution. Within a trial, invoke two prompt-executor agents simultaneously:

Subagent 1:
  agent: prompt-executor
  working_directory: {worktree_original_path}
  expected_base_sha: {pinned_base_sha}
  prompt: {original_request}

Subagent 2:
  agent: prompt-executor
  working_directory: {worktree_optimized_path}
  expected_base_sha: {pinned_base_sha}
  prompt: {prompt_analysis.result.final_prompt}

Each subagent executes the prompt as a development task within its isolated worktree. Clean the pair after collecting both results, then create fresh worktrees for the next trial.

CRITICAL: Both Task tool calls MUST be in the same message to achieve true parallel execution.

Pair validity gate:

• both execution statuses are success;
• both results used fresh worktrees from the same repository state; and
• both results report the pinned base SHA; and
• neither result contains an environment-verification failure.

Keep failed and partial runs as diagnostics only. Continue until three valid pairs are collected or five total trial attempts have run. Compare with reduced confidence when two valid pairs remain. With fewer than two, set status to inconclusive, skip winner/recommendation claims, and report the diagnostics.

Step 4. Environment Cleanup

Execute worktree cleanup per worktree-execution skill "Cleanup" section.

Step 5. Blind Report Generation

Invoke report-generator in two phases.

Phase 1 — blind assessment:

• User task description
• Anonymized valid pairs as Result A and Result B
• No prompts, identity mapping, optimization findings, or change summary

The agent must complete and lock its output-quality judgment before Phase 2.

Phase 2 — identity reveal:

• Identity mapping: A = original, B = optimized
• Full prompt-analysis JSON, not only optimization.finding_resolutions
• Execution metadata and diagnostics for every trial

Output:

• Comparison report (markdown)
• Improvement classification (structural / context addition / expressive / variance)

Quality Gate:

• Output presented to user matches agent's output

The report joins analysis.findings and optimization.finding_resolutions by finding_id; pattern, severity, evidence, change, and source must remain traceable. Context delta is derived from resolutions whose source is a named project path or project knowledge entry.

Step 6. Retrospective

Trigger: Report generation completes

Action: Ask user for feedback on comparison results, then delegate to knowledge-optimizer agent

Improvement Classification

Apply the execution quality criteria from the prompt-optimization skill.

Classification	Definition	Interpretation
Structural	Prompt structure, clarity, specificity improvements	Prompt writing technique
Context Addition	Project-specific information added from codebase investigation	Information advantage
Expressive	Different phrasing, equivalent substance	Neutral
Variance	Within LLM probabilistic variance	Original prompt sufficient

Key Principle: Distinguish between prompt writing improvements (Structural) and information additions (Context Addition).

Final Output to User

Present report-generator's complete output to user. Optimized prompt must appear in full. This is the core learning value of the report.

The report includes (defined in report-generator):

• Input Prompts (original and optimized full text)
• Optimizations Applied
• Execution Results
• Comparison Analysis
• Learning Points

Error Handling

Scenario	Behavior
One side of a trial fails	Exclude the unpaired trial from quality comparison and retain diagnostics
Fewer than two valid pairs	Report inconclusive; no winner or prompt recommendation
All executions fail	Report full failure with diagnostics
Timeout	Terminate, capture partial results, cleanup
Worktree creation fails	Report git error, suggest checking repository state

Prerequisites

• Git repository with git worktree lock support
• Claude Code subagent execution permissions
• Sufficient disk space for worktree copies

Usage Examples

/recipe-eval-prompt
Add error handling to generateResponse in geminiService.ts. Handle 429, timeout, and invalid responses.

/recipe-eval-prompt
Generate code following this skill: .claude/skills/my-skill/SKILL.md

For complex tasks:

/recipe-eval-prompt
Refactor the message pipeline for readability. This may take a while.