BACK TO CLAUDE CODE
Prompting

GPT-5 Prompt Engineering

Guidelines for writing prompts for Alfred's GPT-5 mini/nano models

Prompting Principles

GPT-5 mini and nano have specific prompting characteristics that differ from Claude. These eight principles keep every prompt consistent and efficient.

1

Positive framing over negative

GPT-5 responds better to "do X" than "don't do X". Negative framing wastes reasoning tokens.

Bad

"NEVER generate data from memory. Do NOT fabricate information."

Good

"Use only data from tool results. If a tool fails, say so."

2

No emphasis markers

GPT-5 follows instructions precisely without needing CRITICAL, IMPORTANT, or bold emphasis. These are counter-productive — they burn reasoning tokens trying to assess relative priority.

Bad

"CRITICAL: You MUST call the tools listed above"

Good

"Call the tools listed above to gather data before composing your response."

3

Trust the model’s planning

GPT-5 naturally plans multi-step tool calls. Numbered step-by-step instructions are unnecessary and constrain the model. List what data is needed, not the exact sequence.

Bad

"1. Weather — Use get_weather for London. 2. Calendar — Use list_events with today's date."

Good

"Gather: weather for London (get_weather), today's calendar events, today's tasks."

4

No few-shot examples

GPT-5 performs better with clear descriptive instructions than with input/output examples. Few-shot examples reduce reasoning performance.

Bad

"Instead of 'Task complete' say 'Done. Anything else before I put the kettle on?'"

Good

"Your default is direct and helpful. Dry wit surfaces naturally, not on every message."

5

Resolve contradictions explicitly

GPT-5 burns reasoning tokens trying to reconcile conflicting instructions. If there are exceptions to a rule, state them explicitly as exceptions, not as separate competing rules.

Bad

"Always check both calendars." + later: "If the user says Google, use only Google."

Good

"Check both calendars. Exception: if the user specifies a provider, use only that one."

6

Core instructions first, edge cases after

GPT-5 weights the beginning of the system prompt more heavily. Put identity, behaviour, and core rules at the top. Tool-specific details and edge cases go later.

7

Suppress tool preambles

GPT-5 naturally generates explanatory text before tool calls ("Let me check your calendar..."). On WhatsApp this wastes message space. Include: "When using tools, call them directly without narrating what you're about to do."

8

Use WhatsApp-native formatting

Use the shared WHATSAPP_FORMATTING_RULES constant. WhatsApp supports *bold*, _italic_, and ~strikethrough~ natively. Avoid markdown that WhatsApp cannot render. The sanitizeForWhatsApp() function acts as a safety net.

Prompt Structure Template

When writing a new system prompt, follow this order:

1

Identity \u2014 SOUL.md — who is Alfred, personality

2

Context \u2014 Current time, memory/preferences

3

Behaviour \u2014 Core rules: conciseness, tool usage, memory

4

Formatting \u2014 WhatsApp rules

5

Tool-specific \u2014 Calendar, email, tasks — only when routed

6

Special capabilities \u2014 Voice, etc.

Router Prompt Guidelines

The router uses GPT-5 nano for classification. Keep router prompts:

  • -Under 200 tokens
  • -Structured as a classification task with clear categories
  • -Using response_format: { type: "json_object" } for guaranteed valid JSON
  • -Without synonym lists or keyword matching hints — nano understands intent

Briefing Prompt Guidelines

Briefing prompts should:

  • -List what data to gather (not step-by-step tool sequences)
  • -Include a format template with section headers and placeholder text
  • -Use "Use only data from tool results" instead of anti-hallucination emphasis
  • -Append WHATSAPP_FORMATTING_RULES at the end
  • -State that calendar events and tasks are different things (common confusion point)

Checklist

Before finalising any prompt change:

  • -No CRITICAL/IMPORTANT/NEVER emphasis markers
  • -All instructions use positive framing ("do X" not "don't do X")
  • -No few-shot input/output examples (use descriptive guidance instead)
  • -No numbered step-by-step tool sequences (list what's needed, not the order)
  • -No markdown emphasis used for instruction priority
  • -Core rules appear before edge cases
  • -Tool preamble suppression included for agent-facing prompts
  • -WHATSAPP_FORMATTING_RULES appended (not a custom formatting block)
  • -Contradictions resolved with explicit exceptions
  • -Prompt is concise — every sentence earns its place

Use this as a skill

Copy and add to your Claude Code project as a prompt engineering reference.