---

name: prompting description: "PROACTIVE: This skill SHOULD BE USED AUTOMATICALLY when writing prompts, system prompts, crafting instructions for LLMs, or optimizing AI interactions. Triggers on: write prompt, system prompt, prompt engineering, optimize prompt, better prompt, prompt template, instruction design, context engineering, SKILL.md description, agent instructions. Use for CRAFTING LLM PROMPTS. Model-agnostic best practices for coding agents."

Prompting - Model-Agnostic Best Practices (Late 2025)

Optimal prompt standards for coding agents. Applies to all models (Claude, GPT, Gemini, Grok).

Paradigm: Context Engineering > Prompt Engineering. Focus on designing information systems, not "perfect wording". API-level configs (effort, reasoning_effort, thinking_level) are separate from prompt text.

Quick Reference (Use First)

Before writing/reviewing any prompt, verify:

Copy[ ] Structure: Clear XML/Markdown sections (context → task → constraints → output)
[ ] Order: Long context BEFORE query (+30% quality)
[ ] Task: ONE atomic, explicit action
[ ] Output: Format defined (JSON schema, structure)
[ ] Positive: "Do X" instead of "Don't Y"
[ ] Reasoning: Zero-shot for reasoning models (no manual CoT)
[ ] Examples: Few-shot ONLY for format, not logic
[ ] Caching: Static content first, dynamic last
[ ] Simplicity: Start simple, add complexity only if needed (Stanford 8-word)

Decision Flow

CopyTask Type?
│
├─ Simple (1 step) ────────────► Direct instruction, no CoT
│
├─ Complex reasoning
│   ├─ Reasoning model (o3/o4/GPT-5/Claude thinking)
│   │   └─► Zero-shot + explicit goal only
│   └─ Standard model
│       └─► Structured prompt + optional CoT
│
├─ Code generation ────────────► Define: language, framework, style
│                                Provide: context files, dependencies
│                                Specify: output format, tests
│
└─ Multi-step workflow ────────► Break into atomic tasks
                                 Use Orchestrator-Worker pattern
                                 Include verification steps

Core Structure Template

Copy<context>
  [Background, project details - place LONG documents here]
</context>
<task>
  [ONE specific action required]
</task>
<constraints>
  [Boundaries, limitations]
</constraints>
<output_format>
  [Exact format: JSON schema, markdown structure]
</output_format>

For system prompts (coding agents):

Copy<role>[Expert domain]</role>
<capabilities>[Tools, what agent CAN do]</capabilities>
<constraints>[Security rules, boundaries]</constraints>
<workflow>[Step-by-step process]</workflow>
<output_standards>[Code style, testing requirements]</output_standards>

Key Principles

Technique Selection

See details: references/techniques.md

Context Stack (Priority Order)

1. System Instructions (role, rules) - cached
2. Long-Term Memory (preferences)
3. Retrieved Documents (RAG)
4. Tool Definitions - cached
5. Conversation History
6. Current Task - last for recency

Security

• Separate system/user content with XML tags
• Validate user inputs before embedding
• Validate OUTPUT, not reasoning (CoT hijackable 94-99%)
• Mark sensitive context with explicit boundaries

References

Load as needed for detailed guidance:

• references/techniques.md - AoT, ToT, ReAct, Self-Consistency, Role Reversal, Extended Thinking
• references/anti-patterns.md - Full list of what to avoid with explanations
• references/agentic-patterns.md - Orchestrator-Worker, Reflexion, Plan-Execute, Tool Orchestration
• references/memory-patterns.md - CoALA Framework: Procedural, Semantic, Episodic memory
• references/meta-prompting.md - LLM self-optimization workflow