prompt-improver

name	prompt-improver
description	Improve and optimize prompts for Claude using Anthropic's official best practices and the 5-Element Persona Framework. Use when asked to improve, refine, optimize, or review a prompt. Also triggers on requests to make prompts clearer, more effective, better structured, or to get better/more realistic outputs from Claude. Covers Claude 4.x-specific techniques, persona engineering, constraint-based prompting, and structured outputs.
metadata	{"author":"skeletorjs"}

Prompt Improver

Analyze and improve prompts using Anthropic's official best practices and the 5-Element Persona Framework.

Workflow

Step 1: Identify the Prompt & Intent

• Get the prompt if not provided
• Clarify: Is this for Claude 4.x? API or chat interface?
• Understand: What outcome does the user want to improve?

Step 2: Diagnose the Core Issue

Quick triage against common failure modes:

Symptom	Likely Cause
Generic/shallow output	Missing persona depth or constraints
Wrong format	No output specification or examples
Inconsistent results	Needs few-shot examples with reasoning
Unrealistic suggestions	Missing constraints (budget/time/resources)
Ignores instructions	Needs XML structure or system/user separation
Verbose/over-engineered	Missing explicit direction (Claude 4.x issue)
Off-topic responses	Lacks context about purpose/audience

Step 3: Apply the 5-Element Persona Framework

For any prompt involving expertise or role-play, ensure ALL five elements:

1. Role + Seniority — "Senior Backend Engineer with 8 years..."
2. Industry/Domain Context — "B2B SaaS in fintech, enterprise customers"
3. Methodologies — "Uses JTBD framework, presents at 95% confidence"
4. Constraints — "$50K budget, 6-week timeline, team of 3"
5. Output Format — "2-page executive brief with 3 options + recommendation"

Constraints are the game-changer—without them, Claude gives fantasy answers.

See references/persona-framework.md for templates and examples.

Step 4: Apply Structural Techniques

Select techniques based on the diagnosed issue:

Issue	Technique
Complex reasoning needed	Chain of thought with <thinking> tags
Multiple components	XML tag structure
Format compliance	Prefill (API) + XML output tags
Inconsistent outputs	Few-shot WITH reasoning (input→reasoning→output)
Instruction injection risk	System/User separation

See references/techniques.md for implementation details.

Step 5: Apply Claude 4.x Optimizations

Modern Claude models require explicit direction:

• Be explicit: Add "Go beyond basics" if you want comprehensive output
• Explain the why: Context behind instructions improves compliance
• Tell what TO do: Not "don't use jargon" → "Use plain language a 10th grader understands"
• Match prompt style to output: Reduce markdown in prompt → less markdown in output
• Opus 4.5: Add "Keep solutions simple" to prevent over-engineering

Step 6: Generate the Improved Prompt

Produce the complete improved prompt using this structure:

[SYSTEM PROMPT - when using API or Projects]
You are [role + seniority] with [X years] in [industry/domain].
You specialize in [expertise]. 
Your approach: [methodologies/frameworks]
Your constraints: [budget/time/resources/tradeoffs]
Deliver: [specific output format]

[USER PROMPT]
<context>
[Background, purpose, audience, success criteria, WHY this matters]
</context>

<instructions>
[Clear, numbered steps - tell what TO do, not what NOT to do]
</instructions>

<constraints>
[Negative constraints: "Never assume X", "Do not exceed Y words"]
</constraints>

<examples>
[2-3 examples showing: INPUT → REASONING → OUTPUT]
</examples>

<output_format>
[Exact structure with XML tags for reliable parsing]
</output_format>

[Actual task/content to process]

When to Load References

• references/techniques.md — For detailed technique explanations, API-specific guidance (prefilling), extended thinking tips, or system/user separation patterns
• references/persona-framework.md — When building expert personas, role-based prompts, or adding constraints to get realistic outputs
• references/examples.md — When the user needs before/after demonstrations or few-shot patterns