Custom GPT Prompts: Instruction Templates You Can Reuse

Custom GPT prompts are the instructions you write in the Instructions field when you build a GPT in ChatGPT. A well-structured instruction defines the assistant's role, the tasks it handles, the format it uses for outputs, and any limits on what it should do or say. Tools like fenbs keep your team's approved prompt versions and the lessons each revision taught, so the next person who updates the instructions starts from what is already known.

The templates below follow a four-part pattern: role, tasks, format, and limits. They work as starting points. Each bracketed placeholder marks a decision you need to make for your specific use case. A research assistant and a writing assistant need different role descriptions; the pattern stays the same.

The parts of a Custom GPT instruction

• Role: what the assistant is and what it knows. One or two sentences. Avoid vague labels like "helpful assistant".
• Tasks: the specific jobs it handles. Use a numbered list so you can add or remove tasks without rewriting the whole instruction.
• Format: how outputs should look. Specify structure, length, and any recurring elements like a summary or a caveats section.
• Limits: what the assistant should not do. Explicit limits prevent the assistant from drifting into adjacent tasks you did not intend to support.

Each part does a different job. The role tells the assistant what kind of expertise to draw on. The tasks tell it what actions to take. The format tells it how to structure its answers. The limits tell it where to stop. Instructions that skip one of these parts tend to produce outputs that are inconsistent in the area that was skipped.

Research assistant template

You are a research assistant for [team or company type]. You help users find, summarize, and check information on [topic area].

Tasks:
1. Answer questions about [topic area] based on your training data.
2. Summarize source material the user pastes into the conversation.
3. Flag any claims you are less certain about.
4. Ask for clarification if a question could be interpreted more than one way.

Format: answer first, then reasoning, then caveats. No unnecessary preamble.

Limits: do not draft documents, write marketing copy, or give legal or financial advice. Refer users to a qualified professional for regulated questions.

The "flag uncertain claims" instruction is worth keeping in any research GPT. It produces answers that distinguish between what the assistant knows well and what it is inferring, which makes the output more useful as a starting point for further checking.

Writing assistant template

You are a writing assistant for [role or company type]. You help users draft, revise, and proofread professional documents including [document types: emails, proposals, reports, briefs].

Tasks:
1. When asked to write, confirm the audience, key points, and tone before starting unless all three are already supplied.
2. Produce a draft with clear section headings.
3. After the draft, list two or three suggestions for making it stronger.
4. When asked to revise, identify what changed and what you did not touch.

Format: Markdown unless the user asks for plain text.

Limits: do not add information the user did not supply. If a fact seems missing, ask rather than invent.

The instruction to confirm inputs before writing is a deliberate trade-off. It adds one turn to the conversation but reduces rewrites. Remove it if users are already supplying complete inputs.

Review and feedback assistant template

You are a document reviewer for [team or company type]. You check [document types] against a set of criteria before they are shared externally.

Tasks:
1. When the user shares a document, ask which criteria to check against if they are not already supplied.
2. For each criterion, say whether the document meets it, partially meets it, or fails it, with a one-sentence reason.
3. List any claims that need an external source.
4. Summarize the three most important changes before the document is ready to share.

Format: a table with columns Criterion, Result, Reason; then a numbered list of priority changes.

Limits: do not rewrite the document. Flag issues only.

Keeping the review assistant separate from the writing assistant is a useful design choice. A reviewer that can also rewrite is harder to trust, because it may fix an issue without flagging it. When review and rewrite are separate steps, the person using the output can see what changed and why.

Testing your Custom GPT instructions

Most instruction problems only appear on specific inputs. A GPT that handles the most common task correctly may still fail on edge cases: very short inputs, ambiguous questions, or requests that are slightly outside its defined scope. Testing before sharing the GPT saves the awkwardness of users finding those failures for you.

• Run five representative inputs through the GPT before sharing it. Cover the most common task, an edge case, and an input the GPT should decline.
• Check that limits hold. Ask the GPT to do something outside its scope and confirm it declines clearly rather than attempting a poor answer.
• Test the format. Paste a long and a short input to confirm the output structure stays consistent regardless of input size.
• Ask the GPT to explain its instructions back to you. If it gives a vague summary, your role and task descriptions may need to be more specific.

Lessons from refining instructions

Most Custom GPT instructions improve through revision. A common pattern: the first version is too broad, the second adds too many limits, and the third finds the right scope. Tracking what changed between versions, and why, saves time when you come back to update the instructions months later.

Keep a brief note when you change an instruction: what the problem was, what you changed, and whether it fixed the problem. A shared lessons record in fenbs makes that history available to anyone who maintains the GPT next, rather than leaving it to memory or buried in chat logs.

When a revision solves a recurring problem, record it as a lesson rather than just updating the instructions. The lesson captures why the old version failed. The updated instruction captures what the better version does. Both are useful: the instruction tells the assistant what to do; the lesson tells the next human editor what to preserve.

What to know about Custom GPTs going forward

OpenAI has announced that Custom GPTs will be retired. The Custom GPT guide has the latest details on the timeline and what replaces them. If you are building new GPTs, planning how your approved instructions and lessons will move to whatever replaces them is worth doing now. Storing instructions in a tool outside OpenAI means you have them regardless of platform changes.

For a comparison of Custom GPTs with other ways to configure an AI assistant, Custom GPTs vs Skills covers the main differences. For a broad look at system prompt structure and examples, system prompt examples shows how instructions work across different tools.

How to do this with fenbs

The instructions you write for a Custom GPT are worth keeping outside OpenAI so they survive platform changes and are available to the team without needing to share GPT access.

1. Create a free account. A personal kanban board lets you test the workflow before adding teammates.
2. Go to Skills and prompts in your project. Add each instruction template as its own entry: name, full text, and a note on what the GPT is for.
3. Add a Lessons entry each time you revise an instruction. Note the problem, the change, and whether it worked. The next person who edits the instructions can read the history before making changes.
4. Connect your AI apps at Integrations. A connected assistant can retrieve the current approved instruction text before a task, without you having to paste it manually.