Does the SFT format determine the best prompt?
SFT uses examples of inputs and good responses to shape behavior. Public customer fine-tuning documentation describes messages with user/assistant roles; it does not disclose each ChatGPT model’s internal training examples. Supervised fine-tuning ↗
Behavior and guidance
Some training goals, instruction-following traits and prompting recommendations are documented. Their scope is identified below. A documented trained behavior does not by itself identify a causal contribution from a specific SFT stage.
Internal SFT and a universal optimum
The reviewed documents do not establish complete per-model SFT datasets, mixtures, a fixed Q&A template, or one optimal syntax for every task.
Training records versus a chat request
This authored record illustrates the data structure. It is neither an actual model training sample nor a complete training dataset. In a chat window, describe your task in ordinary language; JSONL wrapping is unnecessary.
{
"messages": [
{
"role": "user",
"content": "Extract the date: The review meeting is on November 12."
},
{
"role": "assistant",
"content": "November 12"
}
]
}Distinguish three kinds of format
- Training record
- An input/target pair for SFT. Customer upload format does not reveal ChatGPT’s internal training records.
- Message role
- API developer/user roles are message metadata. Writing “system:” or XML tags inside a message does not create that role’s authority.
- Content structure
- Headings, tags and examples organize instructions and data. Few-shot prompting supplies examples in context rather than running fine-tuning.
The documentation distinguishes roles, delimiters and few-shot examples. Our conclusion: it does not establish a need to reproduce SFT upload syntax in chat. Prompt engineering · Roles / Formatting / Few-shot ↗
Goal → relevant sources → important constraints → output → completion criteria. For a simple task, include only what helps, in a sentence or two. ChatGPT Learn · Prompting ↗
Different models, different emphasis
Documented guidance comes directly from the sources. Editorial adaptations are starting points based on documented use cases, not measured advantages. Historical guides retain their original model scope. Adapted prompts for four current models were evaluated separately in the pilot above.
ChatGPT · Common
Documented guidanceA starting point for ordinary chat
Official guidance emphasizes useful goals, context and deliverables rather than a required formula.
Apply itStart small for a simple task, then add the constraint that was missed.
Use this starting template →GPT-6 Astra
Documented guidanceModel-specific prompting guidance
The guide documents clarification, formatting and verification behavior, not the underlying SFT examples.
“more likely to ask for clarification”Prompting best practices / Initiative and follow-throughApply itSpecify completion criteria, room for judgment and the desired answer length.
Use this starting template →GPT-6.1 Sol
Editorial adaptationWork and Codex / editorial adaptation
Documented for repeated work across documents, apps and code. No uniquely superior Q&A template was established.
Apply itIdentify source priorities and the final deliverable.
Use this starting template →GPT-6.1 Terra
Editorial adaptationWork and Codex / editorial adaptation
Documented for writing and visual understanding. The wording below is an editorial adaptation for those tasks.
Apply itFor images, specify regions, axes and units; distinguish unreadable content.
Use this starting template →GPT-6 Luna
Editorial adaptationWork and Codex / editorial adaptation
Documented for focused, repeatable tasks such as extraction and classification. A distinct SFT template was not established.
Apply itDefine one task, allowed values and how to handle missing information.
Use this starting template →GPT-4.1
Historical guidanceHistorical model / archived official guide
The guide describes training for more literal instruction following, without disclosing the full SFT recipe.
“more closely and more literally than its predecessors”Introduction / Instruction FollowingApply itTurn implicit expectations into rules and keep examples consistent with instructions.
Use this starting template →o3 / o4-mini
Historical guidanceEarlier reasoning-model guidance
The guide describes training for extended reasoning and advises against requesting internal reasoning traces.
“Keep prompts simple and direct”How to prompt reasoning models effectivelyApply itStart with a clear goal and constraints, without examples. Add consistent examples only when needed.
Use this starting template →Where each model guide applies
The GPT-4.1 guide recommends instructions around long context and says stepwise decomposition may improve quality at a cost in output tokens and latency. These are conditional GPT-4.1 recommendations. GPT-4.1 Prompting Guide ↗
The o-series guide says prompts requesting internal reasoning are unnecessary. Concise calculations, citations and checkable support differ from full internal reasoning traces. Check each model’s documentation and runtime settings for its reasoning capabilities. Reasoning best practices ↗
Is Markdown, XML or JSON best?
Choose a format for its intended use. Use delimiters to separate instructions and source material, make documents easy to read, and structure program outputs to match the required schema. Reasoning best practices ↗
For supported API models, Structured Outputs can enforce a response schema. This differs from asking for JSON in chat, and schema compliance does not establish factual correctness. Structured Outputs ↗
Fill in your task and copy
The selector changes the request guidance only; it does not run a model or switch the model in ChatGPT. Inputs are processed on this page without being saved or sent.
Documented guidance Start small for a simple task, then add the constraint that was missed.
Documented guidance Specify completion criteria, room for judgment and the desired answer length.
Editorial adaptation Identify source priorities and the final deliverable.
Editorial adaptation For images, specify regions, axes and units; distinguish unreadable content.
Editorial adaptation Define one task, allowed values and how to handle missing information.
Historical guidance Turn implicit expectations into rules and keep examples consistent with instructions.
Historical guidance Start with a clear goal and constraints, without examples. Add consistent examples only when needed.
After copying, select the intended model in your own interface. The pilot covers four synthetic Korean tasks with JSON outputs. Check accuracy and revision effort on your own work.
See the example’s grading criteria
Authored reference: A takes 10 + 2×4 = 18 hours; B takes 4 + 4×4 = 20 hours. A saves two hours, but overall superiority remains conditional because features and accuracy are unknown. Grade both the arithmetic and that qualification. This is not a model-generated result or benchmark.
Make the next change specific
In the same conversation, preserve what worked and identify the incorrect fact, new constraint or revision needed. This is an authored follow-up template. ChatGPT Learn · Prompting ↗
Keep: [correct conclusions and constraints]
Change: [incorrect facts or missed requirements]
Additional source: [new evidence]
Return: [changes only / full revision]
After revising, check [specific criteria].How do you find the most efficient format?
Set a quality threshold first, then compare time, cost and revision turns among candidates that pass it. Short answers or fewer tokens alone do not establish a winner. Use task-specific evaluation and human review. Evaluation best practices ↗
- Same tasks and rubric
A small pilot could start with 10–20 representative tasks. This is our experimental suggestion, not a statistically sufficient sample. Score factual errors, omissions and output compliance.
- Change one variable
Hold the model, sources, tools and reasoning settings fixed while comparing a short request with a structured one. Compare models separately using the same request. Repeat cases to record variation.
- Cost of usable results
In ChatGPT, record waiting time, revision turns and whether the result is usable. In API tests, also record input/output and reported reasoning tokens, execution cost and tool costs.
The measured results above include tasks, settings and raw records. Repeat the evaluation when the model version or task changes.
Sources and scope
Official source text was checked on 2026-11-08. Cached search pages contained older model names, so current Markdown sources were cross-checked. The pilot above uses a separate synthetic dataset and fixed scoring criteria.
- Prompt engineering · Roles / Formatting / Few-shot ↗
Separates message-role priority, Markdown/XML boundaries and few-shot examples from fine-tuning.
- ChatGPT Learn · Prompting ↗
General chat guidance on goals, context, output and boundaries; no mandatory syntax.
- Supervised fine-tuning ↗
Customer fine-tuning examples and messages/JSONL format, not disclosure of ChatGPT’s internal training dataset.
- GPT-6 Astra · Prompting best practices ↗
Observed Astra behavior and prompting advice; other GPT-6 models require workload-specific evaluation.
- ChatGPT Learn · Models ↗
Model roles in ChatGPT Work and Codex, distinct from the model picker in regular Chat.
- GPT-4.1 Prompting Guide ↗
An archived guide describing training for more literal instruction following.
- Reasoning best practices ↗
o-series guidance: clear requests, zero-shot first, and no need to demand internal reasoning.
- Structured Outputs ↗
API schema enforcement, distinct from merely asking for JSON in chat.
- Evaluation best practices ↗
Task-specific criteria, representative cases and evaluation calibrated by human review.