Prompt Optimizer

Purpose

Optimize prompts for token efficiency, latency, and output quality while preserving task behavior.

Input: Prompt specification (reviewed), test cases, optimization goal (cost/latency/quality)
Output: Optimized prompt with change log, token comparison, and regression test plan Examples: See examples.md for worked input/output.

Workflow

Step 1: Baseline measurement

Document current prompt:

Step 2: Identify optimization opportunities

Categories:

• Redundancy — repeated instructions
• Verbosity — instructions expressible more concisely
• Example efficiency — fewer/shorter few-shot examples
• Structure — reorder for model attention
• Variable compression — summarize context inputs

Step 3: Apply optimizations

For each change:

Rules:

• Never remove safety constraints for token savings
• Never remove output format contract
• Preserve test case pass behavior

Step 4: Produce optimized prompt

Deliver side-by-side or diff-style comparison:

• Original section → Optimized section
• Net token change

Step 5: Regression test plan

Map each test case from prompt-engineer:

Recommend A/B test if quality risk is Medium+.

Step 6: Validate

Run Validation checklist.

Decision Rules

Validation

• Baseline token metrics documented
• ≥3 optimization changes with rationale
• Optimized prompt preserves output format contract
• Safety constraints unchanged
• Token savings quantified (before/after)
• Regression test plan covers all original test cases
• Changes with quality risk flagged
• No Role Play introduced

Anti-patterns

• Strip safety for tokens — removing constraints to save tokens.
• Optimize without tests — changes without regression verification.
• Over-compress — ambiguous instructions that hurt quality.
• Remove all examples — zero-shot when few-shot was needed.
• False precision — exact token counts without noting estimation uncertainty.

Best Practices

• Optimize iteratively; one change category at a time.
• Use shorter synonyms and bullet lists over prose.
• Move critical instructions to end of prompt (recency).
• Track prompt versions with measured metrics.
• Pair with prompt-reviewer after optimization.

Output Structure

# Prompt Optimization: [Prompt Name]

## Baseline
| Metric | Before |
|--------|--------|
| Template tokens | |
| Total typical | |

## Changes
| # | Change | Tokens saved | Risk |
|---|--------|--------------|------|

## Optimized Prompt
[Full optimized template]

## Comparison
| Metric | Before | After | Delta |
|--------|--------|-------|-------|

## Regression Tests
| Test ID | Must pass | Status |
|---------|-----------|--------|

## Recommendation
Deploy | A/B test first | Revert [change #]

Next Skills