Formulating Effective Requests for External LLM Counsel

Guide for crafting requests that elicit useful, actionable feedback from ChatGPT or Gemini.

Core Principles

External LLMs give better responses when requests are:

• Constrained: A/B choices or ranked lists, not open-ended
• Contextualized: Project goals and priorities stated upfront
• Excellence-framed: Appeal to craftsmanship, not just correctness
• Specific: Concrete code or decisions, not abstract questions

Request Structure

Standard Template

[Context Block - 2-3 sentences]
[Project type and goals. Key priorities or constraints.]

[Question Block]
[The specific decision, with options if applicable]

[Response Format]
Please provide:
- [A/B recommendation with rationale, OR]
- [Ranked list of approaches with tradeoffs]

[Excellence Framing]
Our goal is ἀρετή. Please bring your 職人気質.

Context Block Patterns

Security-focused project:

This is a financial services API where security and auditability are paramount. We prioritize correctness over convenience.

Personal tool:

This is a personal productivity tool I use daily. Speed, trustworthiness, and low friction are my top priorities.

Scalability-focused:

This is a data pipeline expected to handle 10M events/day. Throughput and reliability matter more than development speed.

Startup/iteration:

This is an early-stage product where we need to move fast and learn. Flexibility to pivot matters more than optimization.

Question Block Patterns

Architecture decision (A/B):

We need caching for our API responses. Should we use: A) Redis (external dependency, but battle-tested) B) In-memory LRU cache (simpler, but per-instance)

Our API runs on 3 instances behind a load balancer.

Design pattern choice:

For creating configuration objects with many optional fields, which pattern fits better: A) Builder pattern (fluent, explicit) B) Functional options (Go idiomatic, extensible)

Here's the current struct: [code]

Refactoring approach (ranked list):

This 800-line file needs splitting. What's the best decomposition? [code excerpt showing structure]

Please rank the approaches from most to least recommended.

Code review (ranked issues):

Please review this code and provide a ranked list of corrections and recommendations, from most to least critical. [code]

Response Format Specifications

For A/B decisions:

Please provide:
- Your recommendation (A or B)
- Key factors that determined your choice
- When you'd choose the other option instead

For ranked lists:

Please provide a ranked list from most to least [critical/recommended/important], with brief rationale for each.

For code review:

Please provide a ranked list of issues from most to least critical, covering:
- Correctness and bugs
- Security concerns
- Performance implications
- Design/architecture
- Code clarity

The Excellence Framing

Why It Works

The closing line appeals to craftsmanship values that encourage thoughtful, high-quality responses:

Our goal is ἀρετή. Please bring your 職人気質.

ἀρετή (arete): Greek concept of excellence, virtue, living up to one's full potential

職人気質 (shokunin kishitsu): Japanese craftsman spirit - pride in work, attention to detail, continuous improvement

This framing signals you want thoughtful counsel, not quick answers.

Variations

For different tones:

Formal:

We aim for engineering excellence and long-term maintainability.

Direct:

Give me your honest assessment. I want craft, not compromise.

Collaborative:

Help me think through this carefully. Quality matters more than speed.

Code Review Requests

Scope Limits

• Maximum 10 files per request
• Keep total code under ~2000 lines if possible
• Focus on related files, not scattered snippets

Structure for Code Reviews

[Context]
This is [project type] where [key priorities].

Please review the following code and provide a ranked list of corrections and recommendations.

Our goal is ἀρετή. Please bring your 職人気質.

---

**File 1: path/to/file.go**
```go
[contents]

File 2: path/to/other.go

[contents]

Focus areas:

• Correctness and potential bugs
• Security considerations
• Performance implications
• Design clarity
• Idiomatic usage

### What to Include

- Full file contents (not snippets) when reviewing design
- Relevant type definitions referenced by the code
- Brief note on what the code does if not obvious

### What to Exclude

- Generated code
- Test files (unless reviewing test quality specifically)
- Configuration files (unless that's the focus)
- Dependencies or vendor code

## Anti-Patterns

### DON'T: Open-ended questions

❌ "What do you think about this code?"
❌ "How should I design this system?"
❌ "Any suggestions for improvement?"

### DO: Constrained questions

✅ "Should we use approach A or B? Here are the tradeoffs I see..."
✅ "Rank these three architectural options for our use case..."
✅ "What are the top 3 issues with this code, ranked by severity?"

### DON'T: Missing context

❌ "Review this function: [code]"

### DO: Provide context

✅ "This is a security-critical auth handler. Review for vulnerabilities: [code]"

### DON'T: Too much scope

❌ "Review our entire codebase"
❌ [20 files attached]

### DO: Focused scope

✅ "Review these 3 related handlers for consistency and correctness"

## Service Selection Guidance

**ChatGPT** tends to excel at:
- Implementation details and code generation
- Practical, hands-on advice
- Step-by-step explanations

**Gemini** tends to excel at:
- Architectural reasoning
- Conceptual tradeoff analysis
- Broader design perspectives

Choose based on whether the question is more implementation-focused (ChatGPT) or architecture-focused (Gemini).

## Quick Reference

| Element | Pattern |
|---------|---------|
| Context | 2-3 sentences: project type + priorities |
| Question | Specific decision with options |
| Format | A/B recommendation OR ranked list |
| Closing | Excellence framing (ἀρετή, 職人気質) |
| Code limit | Max 10 files, ~2000 lines |