/check-prompt

Evaluate and improve AI prompt quality

A comprehensive collection of best practices for evaluating and improving the quality of prompts for AI Agents. It systematizes knowledge gained from actual prompt improvement processes, covering all important aspects such as ambiguity elimination, information integration, enforcement enhancement, tracking systems, and continuous improvement.

Usage

# Check the quality of a prompt file
cat your-prompt.md
/check-prompt
"Check the quality of this prompt and suggest improvements"

Options

• None: Analyze current file or selected text
• `--category <name>`: Check only specific category (structure/execution/restrictions/quality/roles/improvement)
• `--score`: Calculate quality score only
• `--fix`: Automatically suggest fixes for detected issues
• `--deep`: Deep analysis mode (focus on ambiguity, information dispersion, and enforcement)

Basic Examples

# Evaluate overall prompt quality
cat devin/playbooks/code-review.md
/check-prompt
"Evaluate this prompt across 6 categories and suggest improvements"

# Deep analysis mode
/check-prompt --deep
"Focus on checking ambiguity, information dispersion, and lack of enforcement to suggest fundamental improvements"

# Check specific category
/check-prompt --category structure
"Check this prompt from the perspective of structure and clarity"

# Detect and fix ambiguous expressions
/check-prompt --fix
"Detect ambiguous expressions and suggest corrections for clarity"

---

Core Design Principles

Principle 1: Completely Eliminate Room for Interpretation

• **Absolutely Prohibited**: "In principle", "Recommended", "If possible", "Depending on the situation", "Use your judgment"
• **Must Use**: "Always", "Absolutely", "Strictly observe", "Without exception", "Mandatory"
• **Exception Conditions**: Strictly limited by numbers ("Only under the following 3 conditions", "Except in these 2 cases")

Principle 2: Strategic Integration of Information

• Completely integrate related important information into one section
• Summarize the overall picture in an execution checklist
• Thoroughly eliminate circular references and dispersion

Principle 3: Building Gradual Enforcement

• Clear hierarchy of 🔴 (Execution stop level) → 🟡 (Quality important) → 🟢 (Recommended items)
• Gradual upgrade from recommended to mandatory level
• Explicit indication of impact and countermeasures for violations

Principle 4: Ensuring Traceability

• All execution results can be recorded and verified
• Technically prevent false reporting
• Objective criteria for success/failure judgment

Principle 5: Feedback-Driven Improvement

• Learn from actual failure cases
• Continuous effectiveness verification
• Automatic detection of new patterns

---

Evaluate and improve AI prompt quality

1. 📐 Structure and Clarity (Weight: 25 points)

1.1 Priority Indication of Instructions (8 points)

• [ ] 🔴🟡🟢 priorities are clearly indicated for all important instructions
• [ ] Conditions for execution stop level are specifically and clearly defined
• [ ] Criteria for each priority level are objective and verifiable
• [ ] Priority hierarchy is consistently applied

1.2 Complete Elimination of Ambiguous Expressions (9 points)

• [ ] **Fatal ambiguous expressions**: 0 instances of "In principle", "Recommended", "If possible"
• [ ] **Use of mandatory expressions**: Appropriate use of "Always", "Absolutely", "Strictly observe", "Without exception"
• [ ] **Numerical limitation of exception conditions**: Clear boundaries like "Only 3 conditions"
• [ ] **Elimination of judgment room**: Use only expressions that cannot be multiple interpreted
• [ ] **Elimination of gray zones**: Clear judgment criteria for all situations

1.3 Strategic Integration of Information (8 points)

• [ ] Multiple location dispersion of important information is completely eliminated
• [ ] Related instructions are logically integrated into one section
• [ ] The overall picture is completely summarized in the execution checklist
• [ ] There are no circular references or infinite loops

2. 🎯 Executability (Weight: 20 points)

2.1 Completeness of Specific Procedures (7 points)

• [ ] All command examples are actually executable and verified
• [ ] Environment variables, prerequisites, and dependencies are clearly stated without omissions
• [ ] Error handling methods are specific and executable
• [ ] The order of procedures is logical and necessary

2.2 Ensuring Verifiability (7 points)

• [ ] Success/failure of execution results can be objectively determined
• [ ] Output examples, log formats, and expected values are specifically shown
• [ ] Testing methods and verification procedures can be implemented
• [ ] Checkpoints for confirming intermediate results are appropriately placed

2.3 Automation Adaptability (6 points)

• [ ] Format that allows easy scripting and CI/CD integration
• [ ] Clear separation between human judgment and AI execution points
• [ ] Support for batch processing and parallel execution

3. 🚫 Clarification of Prohibited Items (Weight: 15 points)

3.1 Systematization of Absolute Prohibitions (8 points)

• [ ] Complete list of operations that must not be performed
• [ ] Explicit indication of impact level (minor/major/fatal) for each prohibited item violation
• [ ] Specific presentation of alternatives and avoidance methods
• [ ] Explanation of technical basis for prohibited items

3.2 Strict Limitation of Exception Conditions (7 points)

• [ ] Conditions allowing exceptions are specific and limited (numerical specification)
• [ ] Objective judgment criteria such as "Completely duplicate", "Explicitly stated"
• [ ] Clear boundaries without leaving gray zones
• [ ] Explicit indication of additional conditions and constraints when applying exceptions

4. 📊 Quality Assurance Mechanisms (Weight: 20 points)

4.1 Completeness of Tracking System (8 points)

• [ ] Automat