ChatGPT Custom Training: Set Up Instructions and Context

To set up ChatGPT for your work, define one task, write its instructions, supply approved reference material, and test the results. This is usually what people need when they ask about ChatGPT custom training; training a model for use by software is a separate process with separate availability constraints.

Use a maintained work record such as fenbs for the rules, decisions, and progress that change between sessions. You can then improve the setup without expecting an old file or a single conversation to hold the current story forever.

How do you build the first working setup?

1. Pick one recurring result, such as a weekly client update, and identify the person who will review it.
2. Write the output requirements and include a short example that the reviewer approves.
3. Collect the sources needed to produce that result. Remove superseded versions and material you are not authorized to share.
4. Place the guidance and references in the appropriate ChatGPT scope, then test the setup in a fresh conversation.
5. Record failures, revise one part, and rerun the same tests before making the setup your normal way of working.

For example, a small design studio might start with proposal follow-ups. The goal is a draft email that preserves the approved scope and asks for one clear next action. “Understand our entire studio” is much harder to test and gives you no clear stopping point.

Make the first success deliberately narrow. A usable follow-up should save review time while preserving facts. You can add another workflow later, with its own requirements, rather than loading every task into one complicated set of instructions.

What should you put in the instructions?

Write what the assistant should do with the material. Include the audience, intended result, required format, and a small number of boundaries. Explain what to do when a required fact is missing. Avoid filling the instruction with company history that never changes the next action.

Draft a follow-up email for the client using the approved scope and current project notes.
Keep scope, dates, and quantities unchanged.
Separate confirmed facts from questions waiting for approval.
Use two short paragraphs and one clear next action.
If the notes do not support a commitment, ask me before including it.
Return a draft for review.

Add an approved example and explain why it works. Perhaps it states the next action clearly without promising an unconfirmed date. That explanation helps you test a new draft on meaning rather than superficial resemblance to one sample.

Use personal settings for preferences you want across your work and project guidance for a particular job. OpenAI’s personalization documentation↗ (opens in a new tab) describes custom instructions as cross-chat preferences. Our custom instructions examples cover writing those preferences; keep this setup focused on the workflow you are building.

Which files and facts should you provide?

Choose sources by whether they answer a question the task must resolve. For the studio example, provide the approved scope, the current update, and the relevant writing sample. A large folder of past proposals may add conflicting dates and obsolete offers without improving this particular email.

• Label the current source and identify the person responsible for maintaining it.
• Separate examples of tone from authoritative facts about the active project.
• Preserve source dates and approval status when those affect what can be said.
• Remove duplicate versions or explain which version replaces which.
• Keep sensitive information out unless your organization has authorized that use.

OpenAI’s projects documentation↗ (opens in a new tab) describes keeping related instructions, files, sources, and chats together. Start the test from the intended project and verify that its required material is available. The projects and custom GPT comparison explains the scope choice; merely naming a document does not establish that the assistant can access it.

Keep fast-changing facts in a record someone can update. “The client has not approved the revised date” belongs with current work, not hidden halfway through a writing sample. fenbs AI context gives a connected assistant the project’s rules, progress, and notes at the start of a session.

How do you test whether the setup is useful?

Prepare three kinds of requests: a normal case, a case missing an important fact, and a case containing an outdated source. Write down the expected behavior before reading the outputs. Otherwise, a fluent answer can persuade you that a missing requirement never mattered.

For the studio, the normal case should produce a clear draft from approved facts. The missing-fact case should surface the unanswered question. The outdated-source case should use the current scope or ask which document controls the answer if you have not established that.

Review the exact failure. Did the assistant lack a source, misunderstand a rule, or apply the wrong example? Fix that cause. Adding a stronger instruction to “be accurate” will not make an inaccessible approval document appear.

Retest the whole small set after a change. Record the instruction version and the sources used so someone else can repeat the comparison. If the setup only works after a long conversation of corrections, the starting guidance still needs work.

When should you consider a separate model project?

Escalate to a technical assessment when you are building software around a stable, repeated task and need to evaluate model behavior systematically. That assessment should include the baseline prompt, representative cases, review requirements, and the operational work of maintaining the application.

Do not assume a fine-tuning service is available because a tutorial recommends it. OpenAI’s fine-tuning availability notice↗ (opens in a new tab) says the self-serve platform is winding down, with access restricted for existing organizations. Check eligibility and the relevant model lifecycle before investing in that route.

The ChatGPT fine-tuning explainer explains the conceptual difference. This practical setup remains useful either way: you still need a clear task, approved examples, current facts, and a way to tell whether a change helped.

How to maintain the setup with fenbs

1. Create a free account and add the project you chose for the first workflow.
2. Save its instruction and approved example in Skills and prompts. State which task the instruction is intended to support.
3. Record approved requirements in Decisions and rules, then add context notes for current facts and source locations.
4. Create a task for testing the setup and save the checks and results there. Add a lesson when a failure changes how future drafts should be prepared.
5. Follow the supported connection guide, ask the assistant to read project context before starting, and save Where we left off after the review.

fenbs does not import your ChatGPT chats or fine-tune a model. It is useful when the workflow needs maintained decisions, tasks, and progress across sessions. For a single private conversation with one file, supplying that file and a clear request may be enough.

Use the integrations guide for your assistant connection and the AI prompts hub for related instruction techniques. Start with the studio’s one follow-up task, verify the result, and keep only the guidance that makes the next review easier.