One-Shot Meta-Prompting Workflow
Build and refine a complete prompt before invoking a slow reasoning model
- Difficulty
- Easy
- Time to result
- ~days to results
- Steps
- 5
- Confidence
- 99%
The One-Shot Meta-Prompting Workflow separates prompt discovery from final reasoning. First, collect the raw context, objectives, constraints, and uncertainties surrounding the task. Discuss them with a fast conversational model that can support multiple inexpensive iterations. Then explicitly ask that model to create a complete prompt for a smarter reasoning model without attempting the final assignment itself. Review the resulting prompt for missing information, conflicting instructions, and an unclear output format. Once the prompt is coherent, submit it as a single context-rich request to the reasoning model. The mechanism avoids waiting through repeated reasoning runs and gives the slower model the information it needs to analyze the problem as a whole.
Origin
Extracted from Marketing Against The Grain, where Sully Omar demonstrates using ChatGPT or Claude to construct prompts that are then submitted to O1 Pro.
Core principles
- 01Reasoning models perform better with complete context
- 02Use fast conversation to clarify an ambiguous request
- 03Have one model draft the prompt for another
- 04Invest preparation time before expensive inference
- 05Explicitly prevent the prompt-writing model from doing the final task
How to run it
- 1
Assemble the raw context
Collect the facts, files, data, constraints, goals, and open questions relevant to the assignment. Completeness matters more than polished formatting at this stage.
Pro tip Copy raw dashboard text or upload supporting material instead of manually retyping it.
Watch out Do not include irrelevant material merely to make the prompt longer.
- 2
Explore with a conversational model
Explain the assignment and use rapid dialogue to clarify what outcome you actually need. Let the model identify missing questions or useful dimensions.
Pro tip Voice conversation can make context gathering faster.
Watch out Do not mistake this exploratory discussion for the final analysis.
- 3
Generate the meta-prompt
Ask the conversational model to turn the discussion into a detailed prompt for a smarter reasoning model. Explicitly instruct it not to complete the final task.
Pro tip Name the desired deliverable, evaluation criteria, and important caveats.
Watch out Without a boundary, the first model may answer the task instead of drafting the prompt.
- 4
Audit the prompt
Check whether the generated prompt contains all relevant context and avoids contradictory goals or metrics. Correct ambiguities before paying for the reasoning run.
Pro tip Ask the conversational model to identify contradictions or missing inputs.
Watch out A reasoning model can analyze bad premises rigorously and still produce the wrong decision.
- 5
One-shot the reasoning model
Send the complete prompt in one submission and allow the reasoning model time to work. Evaluate the result against the stated objective and source facts.
Pro tip If the answer is weak, inspect the input before assuming the model failed.
Watch out Repeated conversational follow-ups can perform worse than a well-formed initial request.
In the wild
A founder dumps raw SaaS metrics into a conversational model and asks it to craft a prompt requesting a detailed plan for improving two or three quarterly metrics. The founder reviews that prompt and submits it to O1 Pro for analysis.
→ The reasoning model returns a structured and prioritized action plan from otherwise messy business data.
A manager discusses a proposed process change with a fast model, including constraints and past failures. The model drafts a single risk-analysis prompt that the manager checks and sends to a reasoning model.
→ The final analysis addresses dependencies, mitigations, and decision criteria without several slow prompt cycles.
Common mistakes
Chatting piecemeal with the reasoning model
Slow reasoning models are poorly suited to many small exploratory turns. Build the request elsewhere and submit it with full context.
Letting the first model do the task
The conversational model may begin the analysis unless explicitly told to create only the downstream prompt.
Skipping contradiction checks
Conflicting metrics or instructions force the reasoning model to guess which premise should control the answer.
Is it for you?
Best for
It is best for difficult tasks where reasoning runs are slow, expensive, or too consequential for loosely improvised prompts.
Not ideal for
It is not ideal for simple questions that benefit more from quick conversational iteration than extensive prompt preparation.
From the transcript
“The right way to use it is sort of like what I like to call as like building context.”
“within prompt engineering, this is called sort of like one-shotting a prompt”
“given this whole chat, please create a prompt that I can give to O1.”
From the episode
Which AI Model Should You Use? (Claude vs GPT & O1 Pro Live Prompt Guide)