prompting-gemini

Critical rules for Gemini 3 Pro that violate typical LLM assumptions.

Non-Negotiable Settings

temperature: 1.0  # NEVER CHANGE - causes loops/degradation at <1.0
top_k: default    # Let model handle
top_p: default    # Let model handle

Unintuitive Strengths to Exploit

Structure Beats Prose

<!-- THIS WORKS -->
<context>
[data here]
</context>
<task>
Analyze the data above
</task>

<!-- NOT THIS -->
Please analyze the following data: [data]

Planning Directives Transform Performance

For ANY complex task, prepend:

Before answering:
1. Decompose the problem into components
2. Create a step-by-step plan
3. Validate your approach
Then execute.

Few-Shot Can Replace Instructions

Instead of explaining what you want, show 2-3 examples. Model infers patterns better than following descriptions.

Context Dumping Works

Put ALL context first, instruction last. Even 50K tokens of context. Model handles it better than interleaved instructions.

Unintuitive Failures to Avoid

Terse by Default

Without explicit verbosity control, expect telegram-style responses.

<constraints>
Verbosity: High  <!-- REQUIRED for detailed answers -->
</constraints>

Politeness Degrades Performance

❌ "Could you please help me understand..."
✅ "Explain X."

Mixed Instructions Cause Chaos

❌ Context, instruction, more context, another instruction
✅ ALL context → ALL instructions

Natural Language < Structured Commands

❌ "Write something creative but not too long"
✅ "Output: 500-word story. Style: surrealist."

Task-Specific Patterns

Complex Analysis

<system>
<reasoning>explicit_chain_of_thought</reasoning>
</system>

<data>[all context]</data>

<analysis_directive>
1. Parse all entities
2. Map relationships  
3. Identify patterns
4. Draw conclusions
Output format: Structured findings with evidence
</analysis_directive>

Creative Generation

<parameters>
Style: [specific aesthetic]
Constraints: [exact limits]
Avoid: [anti-patterns list]
</parameters>

<directive>
First: Commit to ONE bold direction
Then: Execute fully without hedging
</directive>

Code/Technical

<requirements>
- Complete implementation only
- No placeholders/TODOs
- Production-ready
- Include error handling
</requirements>

<task>[specific ask]</task>

Multimodal

[Image/PDF/Media FIRST]

<task>
Describe what you see, then [specific analysis]
</task>

Quick Optimization Checklist

• Temperature = 1.0?
• Used XML/Markdown tags?
• All context BEFORE instructions?
• Explicit verbosity setting?
• Removed all politeness?
• Added planning directive for complex tasks?
• Specified exact output format?
• One clear task, not multiple?
• Examples instead of explanations?

Emergency Fixes

Symptom	Fix
Looping output	Check temperature (must be 1.0)
Too brief	Add <constraints>Verbosity: High</constraints>
Confused execution	Restructure: context→task→output spec
Generic output	Add anti-patterns list + "be distinctive"
Misses instructions	Use XML tags, not prose
Poor reasoning	Add explicit planning phase

The Golden Template

<role>[If needed - domain expert]</role>

<context>
[EVERYTHING relevant - frontload it all]
</context>

<planning>
Before executing, create step-by-step approach
</planning>

<task>
[ONE clear directive]
</task>

<constraints>
- Verbosity: [Low/Medium/High]
- Format: [Exact specification]
- Avoid: [Anti-patterns]
</constraints>

Remember: Gemini 3 is a precision instrument, not a conversation partner. Treat it like a powerful but literal compiler that rewards structure and punishes ambiguity.