Prompt Engineer Pro

Prompt Engineer Pro

Generate and audit production-grade AI agent system prompts using proven architectural patterns.

Workflow

Prompt engineering involves two workflows: Generation (creating new prompts) and Auditing (evaluating existing ones).

Generation Workflow

1. Identify agent type and primary use case
2. Select applicable patterns from the pattern library
3. Compose the prompt using pattern templates
4. Validate against the audit checklist

Auditing Workflow

1. Read the target prompt
2. Run through the audit checklist (see references/audit-checklist.md)
3. Score each section and identify anti-patterns
4. Propose specific improvements with pattern references

Pattern Library

Eight patterns extracted from production agent systems. Read the relevant reference file before using each pattern.

Pattern Selection Guide

Use this to determine which patterns to include based on agent type:

Coding assistant (IDE-embedded, writes/edits code): → Pattern 3 (State Machine) + Pattern 5 (Todo) + Pattern 6 (XML) + Pattern 4 (Scratchpad)

Creative agent (design, writing, presentations): → Pattern 2 (Persona) + Pattern 7 (Design System)

Research/analysis agent (read-only exploration, reports): → Pattern 1 (Skill Injection) + Pattern 3 (State Machine, read-only variant)

Multi-domain agent (handles many task types): → Pattern 1 (Skill Injection) + Pattern 3 (State Machine) + Pattern 5 (Todo)

Full-stack agent (plans, codes, deploys, tests): → All 7 patterns, with Patterns 1-3-5 as core and 4-6-7 as supporting

Prompt Structure Blueprint

Read references/08-prompt-structure-blueprint.md for the full structural analysis of 7 production agent systems (Kimi, Cursor, Claude Code, Devin, Replit, Windsurf).

Universal Section Order

When generating a prompt, follow this canonical order (extracted from all 7 systems):

1. <identity>        — Name, role, 2-3 sentence purpose (ALWAYS first — highest attention)
2. <communication>   — Tone, verbosity, formatting rules
3. <capabilities>    — Explicit CAN/CANNOT lists with boundaries
4. <skills>          — Modular domain knowledge injection (multi-domain agents only)
5. <rules>           — ALWAYS/NEVER imperatives, safety-critical rules first
6. <tools>           — Tool specs with typed params, examples, safety flags
7. <output_format>   — Response structure, code blocks, deliverable format
8. <environment>     — OS, shell, date, workspace paths (ALWAYS last — dynamic)

Three Prompt Archetypes

Select the archetype that matches your agent type:

Attention Optimization

• Identity at the top (highest attention zone)
• Safety rules at the top or bottom (never buried in middle)
• Tool specs can go in the middle (retrieved by name, not position)
• Config/environment at the bottom (dynamic, injected per session)
• Examples near their rules (not grouped separately)

Identity Block Best Practices

• State the name, then the role, then the purpose
• Keep it to 2-3 sentences maximum
• Include the hosting context (IDE name, platform)
• Example: "You are Kiro, an AI assistant and IDE built to assist developers."

Tool Specification Best Practices

For each tool, include:

• Name and one-line description
• Required vs optional parameters with types
• At least one concrete input/output example
• Edge cases and error handling
• Safety flags where applicable (e.g., is_dangerous for shell commands)

Rules Best Practices

• State rules as imperatives: "ALWAYS do X" / "NEVER do Y"
• Group related rules under subheadings
• Prioritize: safety rules first, then behavioral, then stylistic
• Avoid redundancy — each rule should appear exactly once

Validation Scripts

Three Python scripts for automated prompt analysis. Run directly or use within the skill workflow.

Full audit — section coverage, anti-patterns, tool specs, hygiene, scoring (0-10):

python3 scripts/validate_prompt.py <prompt_file> [--format json] [--strict]

Tool spec analysis — extracts tool definitions (XML, JSON, markdown), checks quality:

python3 scripts/analyze_tools.py <prompt_file> [--format json]

Quick lint — fast check with 14 rules, supports multiple files, CI/CD compatible:

python3 scripts/lint_prompt.py <file1> [file2 ...] [--strict]

Audit Quick Reference

Read references/audit-checklist.md for the full checklist. Key items:

Must have: Identity, capability boundaries, tool specs, behavioral rules, safety guardrails Should have: Communication style, error handling, environment context Anti-patterns to flag: Wall of text, vague identity, no boundaries, missing examples, redundant rules Score: 0-3 Poor, 4-6 Fair, 7-8 Good, 9-10 Excellent