Five prompting tricks to get better results from ChatGPT-5

Executive overview

GPT-5 consolidated multiple models into one, using a background router to pick a model, reasoning level, and verbosity for each prompt. Users who prompt poorly get far less value than before — the gap has widened. Five techniques from OpenAI's own research let you steer that router and lift output quality.

The burden of routing is now on the user: better prompts get better models, more reasoning, and more relevant responses.

Trigger words

• Add phrases like "think deeply", "double check your work", "be extremely thorough", or "this is critical to get right" to any prompt.
• These signal the router to assign a higher-capability model and increase reasoning depth.

Prompt optimizer

• OpenAI's prompt optimizer (at platform.openai.com) rewrites a base prompt using best practices.
• It removes vague terms, resolves contradictions, and adds structured checklists the model follows internally.
• Cost is around $1–2 in API credits; output includes a change log explaining every edit.

Words matter

• GPT-5 follows instructions precisely — vagueness and contradictions cause over-reasoning and poor results.
• Bad: "help me plan a nice party, make it fun but not too crazy."
• Good: "Plan a birthday party for an 8-year-old, 10 kids, $200 budget, 2 hours, unicorn theme."
• Specificity eliminates ambiguity before the model starts reasoning.

Structured prompts using XML

• XML tags are OpenAI's recommended structure for system prompts in custom GPTs and projects.
• Wrap content in semantic tags: <context>, <task>, <format>.
• The model comprehends segmented instructions more reliably than unstructured prose.
• If you don't know XML, ask the AI to convert your existing prompt — it handles the formatting.

Self-reflection and rubric iteration

• Instruct the model to generate a rubric aligned with your intent, then judge its own output against it.
• Workflow: create rubric → write first draft → evaluate against rubric (score 1–10, identify gaps) → iterate.
• Typically converges in 2–5 internal iterations; you only receive the final, highest-scoring output.
• Works in direct conversations, custom GPTs, and project system prompts.