Choose a direction — or ask for help choosing

You do not need a goal ready. Think about what you want more of, what you want less of, or something you have been curious about. Learning a language, finishing a creative project, exploring a business idea and working towards a fitness goal are all valid starting points.

If you have several ideas, bring up to three. The builder will compare their meaning, time, cost and trade-offs, then help you choose. If you are still exploring, use a small experiment and a review date instead of forcing a big commitment.

Open the interactive goal builder to fill in what you know and generate a personalised prompt. It works in your browser; your answers are not sent to this website’s server. You can also copy the reusable prompt in step 3 without filling in a form.

2

Open ChatGPT or a dedicated Codex project

ChatGPT: open chatgpt.com, sign in and start a new chat. Paste the builder prompt from the next step, send it and answer the questions. A plain chat is enough to get started.

Codex on desktop: follow the official desktop setup, sign in, choose Codex, create or open a dedicated local project folder and select New chat. Paste the same builder prompt there. If Codex is unavailable in your account, use the ChatGPT route.

In this guide, an agent means a focused assistant with a job, context and a repeatable way of helping. Giving it a role does not grant professional qualifications, tool access, automatic memory or background reminders. Interface labels and availability vary by account and device.

3

Run the agent that helps you build your agent

Copy this prompt into your chosen chat. Tell it what you know so far — or simply say you want help choosing a goal. It should ask a few questions at a time, help you settle on a focus and wait for you to confirm the brief.

The result should be an Agent Pack: a name, complete role instructions, your goal and current-context note, a first-session prompt, a weekly check-in and a simple progress log. Check the goal, starting point, time constraints and assumptions before accepting it.

Your agent builder — copy into ChatGPT or Codex

You are the Be Curious Agent Builder. Help me choose a meaningful goal and create a focused personal AI agent I can use in my own account. Be warm, clear and practical. Explain unfamiliar terms. Ask at most three questions per message and wait for my answers. 1. Discover my outcome, why it matters, starting point, measurable success, deadline (or review date), weekly capacity and non-negotiables. Ask which app and device I use; do not invent interface steps or assume tools are connected. If I am unsure of my goal, start with goal discovery: ask what I want more or less of, what interests me, and what is making life harder. If I have several ideas, compare up to three against meaning, available time, cost and current commitments. Recommend one focus with its trade-off, but let me choose. Help turn the choice into an observable outcome and a realistic review date. Do not choose my goal for me or build an agent before I confirm the focus. A small experiment is appropriate when I am still exploring. 2. Clarify what I want help doing, what I want to keep doing myself, communication style, check-in rhythm and permission boundaries. Flag unrealistic timelines and offer a smaller milestone. For running, ask about recent completed training, typical weekly volume, longest recent run, available days, recovery and relevant injury or health constraints. Do not diagnose, guarantee outcomes or treat an AI role as a professional qualification. Suggest qualified help when the question needs it. 3. Summarise confirmed facts separately from assumptions. Ask me to correct or confirm the brief before producing the final agent. 4. Produce an Agent Pack: name; purpose; complete copyable role instructions; goal and current-context note; first-session prompt; weekly check-in prompt; and a simple progress log. Role instructions must specify responsibilities, boundaries, planning rules, tone, evidence needed for adjustments and what to do after missed actions. Treat scheduled work as planned, never completed. Never invent progress or claim access, memory, reminders or background work that has not been configured. 5. Explain how to use the pack in my actual app: paste into a fresh chat, or use project instructions where available. For a local Codex project, offer a minimal AGENTS.md plus CURRENT.md, but ask before creating or overwriting files. Do not claim installation until verified. Keep scope local and reversible; external writes, calendar edits, messages, purchases and publishing require explicit authorisation. 6. Help me test the agent with my first realistic request and one disrupted-week scenario. Revise unclear instructions. Finish with the next action I can take today. Begin by asking whether I already have a goal or want help choosing one. Follow the relevant path; do not demand a deadline before I know my goal. Do not generate a generic plan before interviewing me.

4

Give your finished agent a home

In ChatGPT: start a fresh chat for your finished agent. Paste the role instructions and current-context note from your Agent Pack together. If you use a project, keep the instructions and context there using the controls available in your account. Keep your own copy of the latest note so you can restart.

In a local Codex project: use the request below after your Agent Pack is ready. AGENTS.md describes the role and working rules; CURRENT.md holds the goal, constraints and latest agreed state. Check the files it creates. Start future chats in the same project and ask it to read both files first.

A saved plan is not evidence of completed work. Ask your agent to keep confirmed progress separate from planned actions and to tell you when context is missing.

5

Create the first plan and test a disrupted week

Ask the agent to restate your goal and constraints, then propose a realistic first week. Each action should have a purpose, a manageable scope and a way to recognise progress. Ask for a smaller version when capacity drops.

Test it with: “My week has changed and I have less time than planned.” A useful assistant asks what happened and chooses a sensible next step. It should not cram everything you missed into the remaining days.

Learning example: an assistant can help define a language milestone, plan practice and review what you recall. Creative example: it can break a draft into steps and give feedback while you keep creative ownership. Business example: it can help identify a question to test and review evidence without assuming success.

Optional running example: someone wants to work towards a first 10 km event and can currently run comfortably for 20 minutes. Their assistant should ask about recent completed training, available days and recovery before deciding whether a timeline is realistic. This is a fictional scenario, not a training prescription. An AI assistant cannot replace qualified coaching or clinical advice; seek individual guidance for relevant health or injury concerns.

6

Make your plan visible and check in weekly

Open the editable weekly planner, enter your reviewed actions, then print or choose Save as PDF. It includes seven days, a weekly rule and a review area. It starts blank and does not prescribe training or choose your goals. Entries are not retained after you reload or close the page.

At each check-in, report what you actually completed, what felt difficult and the time you have next week. Review the proposed changes, then keep an updated context note. Do not assume your assistant can see your calendar, track activity or message you unless those capabilities have been explicitly connected and verified.

7

Keep a reusable copy and adjust what is not helping

You can download the builder instructions and download the weekly check-in. Keep your Agent Pack and latest context note somewhere you can find them.

If the answers are generic, say: “Use my actual baseline and time constraints. Ask what you are missing. Show your assumptions and give me a smaller first week.” Update the role instructions when the same problem repeats.

Keep unnecessary sensitive details out of your prompts. The builder and examples contain no personal training records or private project information. Setup references: official desktop quickstart and projects and chats, checked 18 October 2026. This guide is independent of OpenAI.

Choose a goal and build your own AI agent

Step by step