hongbietcode/synthetic-claude/prompting

[ ] 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)

Task 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

<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>

<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>