Prompt Constraint Analyzer

Analyze prompts using the constraint taxonomy from "How LLMs Follow Instructions: Skillful Coordination, Not a Universal Mechanism" (Rocchetti & Ferrara, Universita degli Studi di Milano, 2026).

Core Research Findings (Your Knowledge Base)

These findings drive ALL analysis decisions:

1. Compositional, not monolithic. LLMs do NOT have a single "instruction-following module." They coordinate separate skills for different constraint types. More types mixed = harder coordination = higher failure risk.

2. Layer stratification. Constraints process at different network depths:

   • Structural (word count, format, JSON) — early layers, fast to detect
   • Lexical (include/exclude words) — middle layers
   • Semantic (topic, sentiment, tone) — late layers, slow to detect
   • Stylistic (register, formality, persona) — late layers

3. Monitoring, not planning. The model does NOT pre-plan constraint satisfaction before generating. It monitors constraints dynamically during token generation. Constraints mentioned earlier in the prompt are monitored longer. Order matters.

4. Asymmetric dependencies. Some skill pairs share representations (topic<->sentiment, exclusion<->toxicity), others are independent. Combining dependent skills is easier than combining independent ones.

5. Model-specific strategies. Claude tends toward constraint-specific encoding (better separation). GPT models vary. Same prompt may need different structure for different models.

Input Format

The skill accepts a file path as argument. The file contains prompts separated by --- on its own line. If there is only one prompt, no separator needed.

Example file:

You are a helpful marketing assistant. Write a 200-word blog post
about AI trends in formal tone. Include the word "innovation" at least twice.
Do not use the word "revolution". Format as 3 paragraphs with headers.
---
Summarize this document in bullet points. Keep it under 100 words.
---
Write a product description for our new feature. Be enthusiastic but professional.

Execution Process

<HARD-GATE> You MUST follow ALL phases in order. Do NOT skip the rewrite phase. Do NOT skip scoring. </HARD-GATE>

Phase 0: Read Input

1. Read the file at the provided path
2. Split into individual prompts by --- separator
3. Number each prompt (P1, P2, P3...)
4. If the file contains structured data (JSON, YAML), extract the prompt/instruction fields intelligently

Phase 1: Constraint Extraction

For EACH prompt, identify every constraint and classify it:

Constraint Types:

Type	Code	What to look for	Examples
Structural	STR	Length limits, format requirements, count requirements, output structure	"max 200 words", "as JSON", "3 paragraphs", "numbered list", "table format"
Lexical	LEX	Required/forbidden words, terminology, specific phrases	"include X", "don't use Y", "start with Z", "mention A and B"
Semantic	SEM	Topic scope, sentiment, meaning boundaries, audience	"about marketing", "positive tone", "for beginners", "B2B focused"
Stylistic	STY	Register, persona, voice, writing style	"formal", "as an expert", "conversational", "academic", "like a friend"

Also detect:

Meta-type	Code	What to look for
Implicit constraint	IMP	Unstated expectations implied by context (e.g., "LinkedIn post" implies hashtags, emoji norms, length norms)
Conflict	CON	Two constraints that pull in opposite directions (e.g., "formal" + "fun and casual")
Dependency	DEP	Constraint B requires constraint A to be satisfied first (e.g., "summarize in formal tone" — must understand content before applying register)
Ambiguity	AMB	Constraint that could be interpreted multiple ways (e.g., "short" without specifying how short)

Output format for each prompt:

## P1: [first 50 chars of prompt...]

### Constraints
1. [STR] "max 200 words" — word count limit
2. [LEX] "include the word 'innovation' at least twice" — term inclusion with count
3. [LEX] "do not use the word 'revolution'" — term exclusion
4. [SEM] "about AI trends" — topic scope
5. [STY] "formal tone" — register
6. [STR] "3 paragraphs with headers" — structure + format

### Meta
- [IMP] Blog post implies intro-body-conclusion flow
- [CON] None detected
- [DEP] Topic (SEM) must be understood before register (STY) can be applied naturally
- [AMB] "headers" — H2? H3? Bold text? Markdown?

Phase 2: Scoring

Calculate three scores for each prompt:

1. Constraint Density Score (CDS)

Count total explicit constraints. Rating:

Count	Rating	Risk
1-2	LOW	Model handles easily
3-4	MEDIUM	Occasional misses possible
5-7	HIGH	Expect partial compliance
8+	CRITICAL	High failure probability, decompose

2. Cross-Type Mixing Index (CTMI)

Count distinct constraint TYPES present (STR/LEX/SEM/STY). This is the key metric from the research — different types activate different network layers with minimal representational sharing.

Types mixed	CTMI	Interpretation
1 type	0.0	Single-skill, high compliance expected
2 types	0.3	Moderate coordination needed
3 types	0.6	Complex coordination, misses likely
4 types	1.0	Maximum coordination load, high risk

Adjust CTMI upward by +0.1 for each:

• Implicit constraint present
• Conflict detected
• Ambiguity detected

Cap at 1.0.

3. Order Risk Score (ORS)

Based on the monitoring-not-planning finding: constraints mentioned earlier are monitored longer during generation. Optimal order places hard-to-satisfy constraints early:

• Semantic/Stylistic constraints SHOULD appear BEFORE structural/lexical ones
• Why: semantic constraints need late-layer processing and longer monitoring. Structural constraints (word count, format) can be checked quickly at the end.

Score:

Order	ORS	Meaning
SEM/STY first, STR/LEX last	OPTIMAL	Matches network processing order
Mixed order	SUBOPTIMAL	Some reordering would help
STR/LEX first, SEM/STY last	POOR	Inverted from optimal; semantic constraints under-monitored

Summary format:

### Scores
| Metric | Value | Rating |
|--------|-------|--------|
| Constraints | 6 | HIGH |
| CTMI | 0.8 (4 types + 1 ambiguity) | HIGH |
| Order | SEM/STY in middle, STR scattered | SUBOPTIMAL |
| Overall Risk | HIGH | Partial compliance expected |

Phase 3: Audit

For each prompt, generate specific findings ranked by severity:

Severity levels:

• CRITICAL — Will likely cause failure. Must fix.
• HIGH — Significant risk. Should fix.
• MEDIUM — Moderate risk. Recommended fix.
• LOW — Minor improvement opportunity.
• INFO — Observation, no action needed.

What to audit:

1. Constraint conflicts (CRITICAL) — two constraints pulling opposite directions
2. Extreme density (CRITICAL) — 8+ constraints, model will drop some
3. All 4 types mixed (HIGH) — maximum coordination load
4. Implicit constraints (HIGH) — model may miss unstated expectations
5. Ambiguous constraints (MEDIUM) — "short", "a few", "some examples"
6. Poor constraint order (MEDIUM) — structural before semantic
7. Redundant constraints (LOW) — saying the same thing twice
8. Missing success criteria (LOW) — no way to verify compliance

Model-specific audit notes:

For Claude:

• Claude handles stylistic constraints well (strong register control)
• Claude tends to be verbose — explicit length limits help
• Claude respects exclusion constraints reliably
• Claude sometimes over-follows structural constraints at expense of semantic quality

For GPT:

• GPT handles semantic constraints well (strong topic adherence)
• GPT sometimes ignores lexical exclusion constraints
• GPT may not follow precise count requirements (e.g., "exactly 3 paragraphs")
• GPT benefits from constraint repetition at the end of the prompt

Audit output format:

### Audit Findings

1. **[HIGH]** CTMI 0.8 — all 4 constraint types present. Model must coordinate
   across all network layers simultaneously. Risk of dropping structural or
   lexical constraints.
   -> Recommendation: Split into 2 prompts or remove least critical constraint type.

2. **[MEDIUM]** Order risk — topic constraint (SEM) appears after format
   constraints (STR). Semantic processing in late layers gets less monitoring time.
   -> Recommendation: Move "about AI trends" to the beginning of the prompt.

3. **[MEDIUM]** Ambiguity in "headers" — unclear format specification.
   -> Recommendation: Specify "## Markdown H2 headers" or "bold text headers".

#### Claude-specific
- Length limit "200 words" is well-placed; Claude tends verbose without it
- Term exclusion "revolution" will be respected reliably

#### GPT-specific
- Consider repeating length limit at the end: "Remember: max 200 words"
- Term exclusion may be ignored; reinforce with "NEVER use the word 'revolution'"

Phase 4: Rewrite

Generate an optimized version of EACH prompt. Create TWO variants:

Variant A: Optimized for Claude Variant B: Optimized for GPT

Rewrite rules (derived from research):

1. Reorder: semantic/stylistic first, structural/lexical last. The model monitors constraints during generation. Place the hardest-to-satisfy (semantic, stylistic) early so they get the longest monitoring window. Place easily-verifiable constraints (word count, format) at the end.

2. Make implicit constraints explicit. If the context implies expectations (e.g., "LinkedIn post" implies certain norms), state them.

3. Resolve conflicts. If two constraints conflict, pick one or add a priority note: "Prioritize X over Y if they conflict."

4. Reduce ambiguity. Replace vague terms with specific ones: "short" -> "under 100 words", "a few" -> "3-5", "some" -> "2-3".

5. Decompose if CTMI > 0.8 and CDS > 7. Split into a chain of 2 prompts. First prompt handles semantic/stylistic constraints. Second prompt applies structural/lexical constraints to the output.

6. Claude-specific adjustments:

   • Be direct with constraints, no need to repeat
   • Use XML tags for structure when mixing many constraints: <constraints>...</constraints>
   • Explicit length limits (Claude is verbose by default)
   • Stylistic instructions work well at the system level

7. GPT-specific adjustments:

   • Repeat critical constraints at the end of the prompt
   • Use numbered lists for constraints (GPT follows numbered instructions well)
   • Bold or CAPS for must-not-violate constraints
   • Lexical exclusions need reinforcement: "NEVER use X. This is critical."

Rewrite output format:

### Rewrite: P1

**Original:**
> [full original prompt]

**Variant A (Claude):**
> [rewritten prompt optimized for Claude]

**Variant B (GPT):**
> [rewritten prompt optimized for GPT]

**Changes made:**
1. Moved topic constraint to beginning (order optimization)
2. Made "headers" explicit as "## Markdown H2" (ambiguity resolution)
3. Added implicit blog structure expectation (implicit -> explicit)
4. [Claude] Added XML constraint block
5. [GPT] Repeated length limit at end, bolded exclusion

Phase 5: Summary Report

After processing all prompts, output a summary table:

## Summary

| # | Prompt (first 40 chars) | CDS | CTMI | ORS | Findings | Verdict |
|---|------------------------|-----|------|-----|----------|---------|
| P1 | Write a 200-word blog post about AI... | HIGH (6) | 0.8 | SUBOPTIMAL | 3 | REWRITE |
| P2 | Summarize this document in bullet... | LOW (2) | 0.3 | OPTIMAL | 0 | OK |
| P3 | Write a product description for... | MED (3) | 0.6 | POOR | 2 | REWRITE |

### Statistics
- Total prompts analyzed: 3
- Need rewrite: 2 (67%)
- Average CTMI: 0.57
- Most common issue: [top finding]

### Top Recommendations
1. [Most impactful recommendation across all prompts]
2. [Second most impactful]
3. [Third most impactful]

Edge Cases

• System prompts: Treat as a single prompt. System prompts often have high CTMI by nature — flag but don't necessarily recommend decomposition.
• Few-shot prompts: Analyze the instruction part, not the examples. Note if examples contradict instructions.
• Chain-of-thought prompts: CoT instructions ("think step by step") are META constraints — note them but don't count toward CTMI.
• Tool-use prompts: Tool definitions are STRUCTURAL. Function calling instructions are STRUCTURAL. The task description follows normal analysis.
• Empty or trivial prompts: If a prompt has 0-1 constraints, mark as OK — no optimization needed and skip rewrite.

What This Skill Does NOT Do

• Does NOT evaluate prompt quality for a specific domain (marketing, coding, etc.)
• Does NOT test prompts against actual models
• Does NOT guarantee compliance improvement — it predicts risk based on research
• Does NOT replace human judgment on prompt intent