Meta-Prompting
A prompt that makes your prompts better. Nate B Jones’s answer to the GPT-5 prompting problem: instead of trying to write a perfect detailed prompt yourself, you give the model a structured instruction-set for how to interpret, restructure, and execute whatever sloppy request you actually type. Meta-prompting acts as power steering for a model that Nate describes as “a speedboat with a really big rudder — it wants to go fast and it wants to be steered really hard.”
Why GPT-5 Specifically Needs This
GPT-5 changed the prompting game in four ways that break old habits:
- It’s multiple models behind a router. The way you structure the prompt influences which underlying model GPT-5 calls. Bad structure → wrong model gets the job.
- The Precision Tax. Contradictory instructions (“be brief but comprehensive”) force the model to burn tokens trying to resolve the conflict, and the output suffers.
- Agentic bias. GPT-5 doesn’t want to chat — it wants to complete missions. Casual conversational prompts make it fabricate details to fill the mission shape.
- The Expertise Paradox. It’s marketed to non-experts but works best with expert-structured prompts. Anthropic released a GPT-5 prompting guide because OpenAI internally recognized this gap.
“The era of casual conversation prompting is over. With ChatGPT-5, we need systematic prompting.” — Nate B Jones
A Quick Example Meta-Prompt
The “five-minute get started” meta-prompt Nate uses:
Transform my request into a structured brief and then execute it.
First, interpret what I’m actually asking for: what type of output would help me, what expertise would be relevant, what format would be useful, what level of detail.
Then restructure and execute as: a specific role (you should infer appropriate expertise), a specific objective (please make my vague request more specific), an approach (choose the methodology that fits the objective), and an output.
When Nate gave ChatGPT-5 the bare prompt “help me prepare for tomorrow’s meeting,” it fabricated an agenda, a 30-minute timeline, fake industry stats — useless. With the meta-prompt wrapper around the same five words, the model verbalized assumptions it could be corrected on, asked precise clarifying questions, and produced a meeting prep sheet at “80% good” on the first pass.
The Seven Prompting Principles for GPT-5
- Structure beats intelligence. Headers, bullets, methodologies, and explicit format expectations steer the implicit router toward the right underlying model. “Think hard” works too, but structure is more reliable.
- Explicitly prioritize tension. “My primary goal is X. My secondary goal is Y. When in doubt, prioritize one over two.” Resolves the Precision Tax.
- Depth ≠ length. GPT-5 separates reasoning depth from response verbosity. Tell it both: “I want PhD-level analysis in a tight executive summary.”
- Define uncertainty protocols. GPT-5 will literally attempt any task. You must tell it what to do when data is missing or ambiguous: “If you can’t find X, ask. Don’t fabricate.”
- Be opinionated about tool use. GPT-5 is either a tool maximalist or minimalist. Specify: “First search the web. Then analyze the data this way.”
- Context memory is an illusion. It’s re-reading everything each turn. Periodically restate instructions. The “flag” trick: tell it to write
flagat the end of every response while it remembers the instruction — when the flag disappears, you know it forgot. - Give the model methodologies, not just goals. A clear stepwise process beats “think harder.”
The Seven Components of a Good Prompt
If you’re writing from scratch instead of using a meta-prompt:
| # | Component | Why |
|---|---|---|
| 1 | Role | Aims the implicit router toward the right expertise tier |
| 2 | Objective | GPT-5 needs missions, not conversations |
| 3 | Process methodology | Step-by-step path to the result |
| 4 | Format expectations | The model wants to deliver — tell it the shape |
| 5 | Boundaries / anti-goals | ”Don’t go to the coral reefs” — where not to steer |
| 6 | Uncertainty protocols | Explicit priorities when there’s ambiguity |
| 7 | Validation criteria | How the model can check its own work |
When NOT to Meta-Prompt
Skip the heavy machinery for:
- Simple factual queries
- Exploratory or discovery conversations
- Personal/emotional conversations (Nate’s note: “Claude has better emotional capabilities than ChatGPT — pick the right model”)
Meta-prompting is built for mission-shaped work. On the edges of what GPT-5 does well, drop the structure or use a different model.
Relationship to Other Wiki Concepts
- four-prompting-disciplines — Nate’s broader framework. Meta-prompting sits at the intersection of context engineering and specification engineering: you’re giving the model both the right context-shaping instructions AND a spec for the output structure.
- ai-coding-workflow — Cole Medin’s PLANNING.md/TASK.md pattern is meta-prompting for coding agents: you’re not writing a perfect prompt every time, you’re encoding the discipline once in a markdown file the agent reads.
- bmad-method — The BMAD persona artifacts are essentially heavy meta-prompts: each role’s output is the next role’s prompt.
- ChatGPT — the model family meta-prompting was designed for.
See Also
- nate-b-jones — author
- chatgpt — the target model family
- four-prompting-disciplines — the broader prompting taxonomy
- ai-coding-workflow, bmad-method — applied meta-prompting in coding contexts
- Source: ChatGPT-5 Prompting