Optimize Prompt

Pipeline

P → [Input Analysis] → [Critique] → P' → [Equiv Check 1] → [Compress] → P'' → [Equiv Check 2] → Output P''

Stage 0: Input Analysis

1. Does P take input? Check for placeholders (

   {{X}}

   ,

   {input}

   ,

   [USER INPUT]

   ), references to "the user's input"/"the given text"/"the provided data", or expectation of concatenation with external content.
2. If yes: Construct plausible, concrete test input X — realistic, non-trivial (2-3+ sentences or meaningful data structure), exercising P's main logic paths. If P has branching conditions, X triggers the primary branch. If P produces structured output, X requires all output fields. If P has validation conditions, X passes validation. State X explicitly.
3. If no: P is self-contained. Set X = ∅. Equivalence checks compare outputs of P directly.

**Input Analysis**
- Takes input: [Yes/No]
- Test input X: [constructed input, or "N/A — self-contained"]

Stage 1: Critique

1. Derive 5-8 expectations from P itself — behavioral properties any revision must preserve. Frame as testable statements (e.g., "Produces JSON output", "Rejects off-topic queries").
2. Evaluate P against expectations. Per expectation: confidence (0-100%) with concise rationale referencing specific text in P.
3. Propose fixes for expectations with confidence < 95%. Each fix: exact text to add, remove, or replace.
4. Produce P' — revised prompt with all fixes applied. Complete text, not a diff.

**Stage 1: Critique**

| Expectation | Confidence | Rationale |
|---|---|---|
| ... | ...% | ... |

**Overall Score: [average]%**

**Fixes Applied:**
1. [Fix with exact text changes]
...

---
> **P' (Revised Prompt):**
---
[Full text of P']
---

Stage 2: Equivalence Check 1

1. Simulate running P on X. Describe expected output (2-4 sentences): structure, content, tone, key features.
2. Simulate running P' on X. Describe expected output (2-4 sentences).
3. Equivalent if all hold:
   • (a) Output structure identical (sections, format, ordering)
   • (b) Factual/decisional content identical — nothing added, removed, or altered
   • (c) Differences limited to: wording improvements, added specificity, stronger constraint enforcement
   • (d) User expecting P's behavior would accept output without noticing intent change
   • If (a)-(c) hold but (d) uncertain, default YES.

**Stage 2: Equivalence Check 1**

- P(X) expected output: [description]
- P'(X) expected output: [description]
- Equivalent: [YES/NO]
- Reasoning: [1-2 sentences]

**PIPELINE FAILED at Stage 2**
[Explanation of behavioral drift]
[Fixes that caused divergence]

Stage 3: Compress

1. Target 10-30% token reduction, 100% semantic retention.
2. Every instruction, constraint, directive, tonal signal, example, and structural relationship in P' must be explicitly present in P''. Nothing left to inference.
3. Allowed: remove filler, collapse redundancy, tighten syntax, merge duplicates, normalize structure.
4. Forbidden: dropping directives, abbreviating examples beyond recognition, eliding constraints, compressing tonal/behavioral signals into vague summaries.
5. Produce P'' and a directive map (each P' directive → P'' counterpart).

**Stage 3: Compress**

---
> **P'' (Compressed Prompt):**
---
[Full text of P'']
---

**Directive Map:**
| # | Original directive (P') | Compressed counterpart (P'') |
|---|---|---|
| 1 | [directive from P'] | [text in P''] |
| ... | ... | ... |

**Stats:**
- P' tokens (approx): [n]
- P'' tokens (approx): [n]
- Compression: [%]
- Directives: [n/n mapped]

Stage 4: Equivalence Check 2

**Stage 4: Equivalence Check 2**

- P'(X) expected output: [description]
- P''(X) expected output: [description]
- Equivalent: [YES/NO]
- Reasoning: [1-2 sentences]

**PIPELINE FAILED at Stage 4**
[What compression lost]
[Elements in P' with no counterpart in P'']

Final Output

**PIPELINE SUCCEEDED**

Your optimized prompt (P''):

---
> **BEGIN OPTIMIZED PROMPT**
---
[Full text of P'']
---
> **END OPTIMIZED PROMPT**
---

**Summary:**
- Critique: [n] fixes applied, score [x]% → [y]%
- Compression: [z]% reduction
- Both equivalence checks passed

Edge Cases

• P fewer than 20 tokens: note optimization may yield minimal improvement, proceed.
• P already optimal (all expectations >= 95%, compression < 10%): state "Prompt is already well-optimized", return P unchanged.
• Adversarial or self-referential P: evaluate literally, note observation.
• Compression < 10% reduction without dropping directives: skip compression, set P'' = P', note "Compression skipped — prompt already dense."
• P' more than 50% longer than P: flag "Significant expansion", verify in Stage 2 that expansion only adds guardrails/specificity.
• P contains code blocks: preserve verbatim during critique and compression unless fix explicitly targets code content.

Input