AI For Dev

summary-nate-b-jones-chatgpt5-prompting

3 min read

ChatGPT-5 Prompting is Too Hard: This Video Makes it Easy for You

Channel: Nate B Jones Format: YouTube video Published: 2025-09-04 Sponsor: None disclosed (links to Nate’s own Substack and site)

Summary

Nate B Jones explains why GPT-5 prompting feels harder than earlier models and introduces meta-prompting as the practical fix. GPT-5 is “a speedboat with a really big rudder” — extremely powerful, agentic, literal, and easy to misguide. Without structure, it fabricates details to fill the mission shape you implied. Meta-prompts wrap your sloppy human request in an instruction-set that tells the model how to interpret, structure, and execute it. The video walks through a live before-and-after on “help me prepare for tomorrow’s meeting” — useless without a meta-prompt, 80% useful with one. See meta-prompting for the full concept page.

Key Points

  • GPT-5 is a different prompting beast than prior models: a router over multiple sub-models, agentic by default, literal in execution, demands precision
  • The four GPT-5 differences: (1) multiple models behind a dispatcher; (2) the Precision Tax — contradictory instructions burn tokens; (3) agentic bias — wants missions, not conversations; (4) expertise paradox — marketed to non-experts but needs expert-structured prompts
  • Live demo: bare “help me prepare for tomorrow’s meeting” prompt → fabricated agenda, fake stats, useless. Same prompt wrapped in a meta-prompt → model verbalized assumptions, asked precise questions, produced an 80%-good first-pass meeting brief.
  • Meta-prompting = power steering — instead of writing a perfect prompt, you give the model a structured wrapper that tells it how to interpret and execute whatever you actually typed
  • Seven prompting principles for GPT-5: structure beats intelligence; explicitly prioritize tension; depth ≠ length; define uncertainty protocols; be opinionated about tool use; context memory is an illusion (use the “flag” trick to detect forgetting); give methodologies, not goals
  • Seven components of a good GPT-5 prompt: role, objective, process methodology, format expectations, boundaries/anti-goals, uncertainty protocols, validation criteria
  • Don’t over-engineer: meta-prompts are for mission-shaped work. Skip them for simple factual queries, exploratory conversations, or personal/emotional ones — and use Claude for emotional conversations because “Claude has better emotional capabilities than ChatGPT, just say it”
  • The flag trick: tell GPT-5 to write flag at the end of every response while it remembers the initial instruction. When the flag disappears, you know it dropped context.
  • The shift is systematic prompting > casual prompting — Nate expects GPT-6 to demand even more precision

Notable Quotes

“GPT-5 is a speedboat with a really big rudder — it wants to go fast and it wants to be steered really hard.” — Nate B Jones

“The era of casual conversation prompting is over. With ChatGPT-5, we need systematic prompting.” — Nate B Jones

“Structure beats intelligence — give the model methodologies, not just vague instructions.” — Nate B Jones

Sponsorship & Bias Notes

Sponsor: None disclosed. Nate links to his own Substack (“the ChatGPT-5 prompting manual”) and natebjones.com — self-promotion of his own deeper resources, not third-party paid content.

Product placement / affiliations: The video is essentially a soft launch for Nate’s own paid Substack content with the full meta-prompt library. Disclosed openly.

Comparison bias: Nate explicitly claims “Claude has better emotional capabilities than ChatGPT” — not benchmarked, but consistent with his broader editorial positioning across other videos.

Connected Pages

See Also

Graph View

Backlinks