ChatGPT Agent Prompts: Templates With Clear Checkpoints

A ChatGPT agent prompt is the instruction set that tells the agent what goal to pursue, what it may do to reach that goal, when to stop and check with you, and how to record what it did. Vague prompts produce unpredictable results. Clear prompts with explicit checkpoints give the agent permission to work while keeping you in control of the decisions.

The templates below cover the most common agent tasks. Each one is designed to be copied, filled in and saved as a reusable starting point. fenbs keeps your approved prompt versions and project rules in a shared place that any connected AI app can read before starting.

The five parts every agent prompt needs

Before looking at specific templates, it helps to understand the components that every reliable agent prompt shares. Leaving any of these out is usually where problems start.

• Goal: one sentence stating what a successful run produces. Vague goals produce vague outputs.
• Inputs: what the agent has access to (files, web search, tools, data sources) and what it must not touch.
• Permissions: what actions it may take on its own versus what requires your approval first.
• Checkpoints: explicit points where the agent stops, shows you what it has done so far, and waits for your go-ahead before continuing.
• Output format: how the agent should present its results, including how to flag uncertainty.

Template: research and summarize

Use this template when you want the agent to gather information and produce a structured summary. Fill in the bracketed sections before using.

Goal: Produce a structured summary of [topic], covering [specific angle or question].

Inputs: You may search the web. Do not access [excluded sources or types].

Permissions: You may conduct searches and read linked pages without checking with me first.
Do not draft or send anything on my behalf.

Checkpoint 1: After gathering at least five sources, show me your source list and the main claim from each. Wait for my confirmation before drafting the summary.

Output: A summary of [length or format]. For each key claim, note the source. Flag anything you could not verify.

Template: draft and revise

Use this template when the agent will produce a written draft based on inputs you provide. The checkpoint before final delivery is the most important part here.

Goal: Write a [document type] that [purpose and target audience].

Inputs: Use only the materials I provide in this session. Do not search the web for additional sources.

Permissions: You may write a full draft without checking at each paragraph.
Do not share or publish the draft in any way.

Checkpoint 1: Before writing, outline the structure in three to five bullet points. Wait for my approval of the structure before drafting.

Checkpoint 2: When the draft is complete, show me the full text before applying any further revisions.

Output: A draft in [format]. Mark any sections where you made assumptions about tone or detail.

Template: data review and report

Use this template when the agent will read a file or dataset and produce a report. Keep the permissions narrow here: data review agents that can also send files create risk.

Goal: Review [file or data source] and produce a report covering [specific questions or metrics].

Inputs: The file I attach to this session. Do not access other files or external data.

Permissions: Read and analyze the attached file only. Do not modify, copy or transmit the file.

Checkpoint 1: After an initial scan, tell me what the file contains and note any potential data quality issues. Wait for my confirmation before proceeding.

Output: A report in [format] answering the specified questions. For each finding, state where in the data it comes from. Note clearly what you could not determine.

Template: task list and plan

Use this template when you want the agent to break a project into tasks and produce a plan. This is a planning prompt, not an execution prompt: the agent proposes; you decide what to act on.

Goal: Create a task list and rough plan for [project description].

Inputs: The context I provide below. Do not search for additional information unless I ask.

Permissions: You may draft a full task list and plan without interrupting for each step.
All decisions about priority, scope or resources stay with me.

Checkpoint 1: Before detailing the plan, list the five to ten main tasks. Wait for my review before adding subtasks or timeline estimates.

Output: A task list in [format] with a brief rationale for each item and the main dependencies between tasks. Flag any task where you lack enough information to estimate scope.

For related prompt templates, the system prompt examples guide covers standing instruction patterns. The chatgpt prompts for work guide has shorter templates for single-turn tasks.

Storing and improving your agent prompts

A prompt that worked well once is worth saving. A prompt that failed tells you something useful about what to add or remove. Treat your agent prompt library as a living set of instructions that gets better each time you run and review it.

When you revise a prompt, keep a note of what you changed and why. If a checkpoint was too early or too late, record that. If the output format was wrong, specify it more precisely. A well-maintained library of agent prompts saves setup time and keeps your work consistent across sessions and models.

For running longer agent tasks, the chatgpt agent mode guide explains how ChatGPT's built-in agent mode works and what it can do. The hermes agent guide looks at a different agent framework for comparison.

When to revise an agent prompt

A prompt that worked last month may fail today because the task changed, the data format changed, or the AI app updated its behavior. Two signals tell you a prompt needs revision: the output is consistently wrong in the same way, or you are manually fixing the same thing after every run.

When either signal appears, trace it to a specific part of the prompt. Is the goal unclear? Is a checkpoint missing? Is the output format too loose? Revise that section and test again before updating any saved version. Changing a prompt and saving it before testing often creates a worse version than the one it replaced.

Keep a short note alongside each saved prompt: what it is for, what problems it solved, and the date of the last revision. A prompt without any history is easy to misuse or accidentally improve in the wrong direction.

How to do this with fenbs

fenbs keeps your approved agent prompts in a shared Skills and prompts library that any connected AI can read. Project rules and decisions are available in the same place, so the agent starts with both the right prompt and the current project context.

1. Create a free account. Set up a project for the work these agent prompts belong to.
2. Add your approved prompt templates as Skills in the project. Give each one a clear name and a note on what task it suits.
3. Connect ChatGPT or another supported AI app over MCP at Integrations. The app can list your skills and read the full prompt before starting a session.
4. Write the project's standing rules as Decisions and rules. These appear first in the AI context, before the skill, so the agent knows what is always off-limits.
5. After each session, note what the prompt produced and update the skill if you improved the wording. The revision history stays in fenbs.