prompt-crafter

name: prompt-crafter description: Create, optimize, critique, and programmatically structure prompts for AI systems. Use this skill whenever the user is designing or improving a static prompt, system prompt, coding prompt, agent prompt, workflow prompt, MCP-oriented prompt package, or an algorithmic prompt optimization pipeline. Also use it when the user asks to turn vague AI behavior into a precise instruction set, tool policy, agent spec, evaluation metric, or prompt architecture.

Prompt Crafter

Prompt Crafter is an AI prompt and agent design framework.

Use it to turn loose intent into a production-ready artifact for prompt engineering, agent design, workflow design, and MCP-oriented prompt packaging.

Core Principle

Do not ask a fixed intake questionnaire.

Instead:

1. Detect what the user is actually trying to produce.
2. Collect only missing critical context.
3. Calibrate for the target model or tool when relevant.
4. Return a structured artifact, not vague advice.
5. Critique and improve the result when the user asks for optimization or evaluation.

Supported Request Types

Detect the closest match before producing anything:

• prompt: a general task prompt for an LLM
• system_prompt: system or instruction-layer prompt
• agent_prompt: prompt plus operating rules for an agent
• workflow_prompt: multi-step prompt or orchestration flow
• coding_prompt: prompt tuned for coding agents and IDE assistants
• image_prompt: prompt for image generation systems
• evaluation_prompt: rubric, grading prompt, or critique harness
• mcp_prompt_package: prompt package for MCP-style tool usage, tool descriptions, config scaffolds, and approval rules
• programmatic_prompt: algorithmic prompt pipeline design (Task Signature, Evaluation Metric, Dataset Blueprint, Optimizer Strategy)

If the user request spans multiple types, choose the primary one and mention the secondary ones in the output.

Discovery Rules

Ask questions only when the missing information would materially change the artifact.

Typical missing context:

• target model or tool
• target audience or end user
• desired output format
• constraints or non-goals
• available tools or data sources
• risk tolerance for automation

Skip questions when the user already provided enough context to produce a strong first draft.

If questions are needed:

• ask only the minimum number
• prefer high-impact questions
• explain the reason briefly
• proceed as soon as the critical gap is closed

Routing

After request-type detection, load only the reference files that matter:

• model or tool calibration: references/tool-profiles.md
• prompt structures: references/prompt-patterns.md
• agent specs: references/agent-templates.md
• workflow specs: references/workflow-templates.md
• prompt review or grading: references/evaluation-rubrics.md
• retrieval grounding: references/rag-patterns.md
• MCP-oriented packaging: references/mcp-templates.md
• multi-agent coordination: references/multi-agent-patterns.md
• examples and few-shot inspiration: references/examples.md
• fast preflight and QA: references/checklists.md

Do not load every reference file by default.

Output Contracts

Always return an artifact that matches the request type.

prompt

Return:

1. Goal
2. Final Prompt
3. Why This Structure Works
4. Optional Tweaks

system_prompt

Return:

1. Role
2. Behavior Rules
3. Constraints
4. Output Contract
5. Final System Prompt

agent_prompt

Return:

1. Agent Role
2. Inputs
3. Tools
4. Memory Policy
5. Decision Rules
6. Escalation Rules
7. Output Contract
8. Final Agent Prompt

workflow_prompt

Return:

1. Workflow Goal
2. Stages
3. Stage Instructions
4. Quality Gates
5. Final Workflow Prompt

coding_prompt

Return:

1. Task Summary
2. Constraints
3. Validation Expectations
4. Final Coding Prompt

image_prompt

Return:

1. Subject
2. Composition
3. Style
4. Negative Constraints
5. Final Image Prompt

evaluation_prompt

Return:

1. Artifact Under Review
2. Criteria
3. Scoring Logic
4. Failure Conditions
5. Final Evaluation Prompt

mcp_prompt_package

Return:

1. Package Goal
2. System Prompt Guidance
3. Tool Description Guidance
4. Approval and Escalation Policy
5. Workflow Rules
6. Evaluation Checks
7. Config Scaffold

For V1, this package is documentation and scaffolding only. Do not imply executable correctness for any specific MCP runtime unless the user supplied that runtime and asked for a concrete adapter.

programmatic_prompt

Return:

1. Task Signature: High-level description of the task, input fields, and output fields.
2. Evaluation Metric: Scoring logic or 'LLM-as-a-judge' criteria for programmatic evaluation.
3. Dataset Blueprint: Structure of the examples needed to optimize the prompt (train/dev/test splits).
4. Optimizer Strategy: Recommended algorithmic optimization approach (e.g., automated few-shot selection, iterative prompt tuning) based on the task.

Optimization Mode

When the user asks to improve an existing prompt, follow this sequence:

1. Identify the request type.
2. Diagnose the current weakness (e.g., manual eye-balling vs programmatic scoring).
3. If the user is building a production LLM application, explicitly recommend transitioning to a programmatic_prompt approach (defining metrics and datasets for algorithmic optimization) rather than manually tweaking static words.
4. Keep what already works.
5. Improve structure, specificity, and tool alignment.
6. Return the revised artifact plus the rationale.

Common improvements:

• clearer role framing
• stronger output contract
• reduced ambiguity
• tighter tool usage rules
• better retrieval grounding
• stronger evaluation criteria
• cleaner escalation boundaries

Evaluation Mode

When the user asks to evaluate or critique a prompt:

1. State the detected request type.
2. Score it using the most relevant rubric from references/evaluation-rubrics.md.
3. Identify the top failure modes.
4. Provide a revised version if useful.

Focus on:

• clarity
• completeness
• model or tool alignment
• robustness to ambiguity
• safety for tool-enabled systems

Safety Rules

Do not produce prompts or agent packages that are designed to:

• exfiltrate secrets or credentials
• bypass platform safeguards
• automate unauthorized access
• hide dangerous tool actions from humans
• weaken approval boundaries for high-risk actions
• facilitate prompt injection abuse or data leakage

For agent, workflow, and MCP-oriented outputs, default to:

• least-privilege tool framing
• explicit approval gates for destructive or external actions
• clear fallback behavior when inputs are incomplete or outputs are invalid
• auditable output contracts

Trigger Boundaries

This skill should trigger for:

• prompt engineering
• prompt optimization
• system prompt design
• coding prompt design
• prompt critique
• agent design
• workflow design
• MCP-oriented prompt and tool-policy design
• algorithmic prompt optimization (evaluation metrics, dataset blueprints, automated tuning)

This skill should not trigger for:

• generic essay writing
• generic copyediting
• general software architecture not centered on AI prompting or agent behavior
• unrelated design tasks with no prompt or agent component

Operating Notes

• Prefer principle-based guidance over provider-specific trivia.
• If the user names a target model, calibrate to it.
• If no model is named, produce a neutral artifact that is portable across modern LLMs.
• Keep outputs implementation-ready.
• Avoid fluff and avoid meta commentary about prompting.

Final Check

Before finishing, verify:

1. The request type is correct.
2. The artifact matches the output contract.
3. Critical missing context was either collected or consciously assumed.
4. Tool and safety rules are explicit when relevant.
5. The result is ready to use without a second rewrite.