Go beyond
asking questions.

Start with the kind of task you want done. You do not need every tool to begin.

• ChatGPT: ask questions, explain ideas, and draft text.
• Work: ask ChatGPT to complete a task using files and tools.
• Codex: work on a software project, edit its files, and run tests.

Where should it run?

How to tell: in desktop Work, check Work locally / Cloud. In Codex, look for Local / Worktree / Cloud. Worktree is local, too. Controlling a computer from your phone does not turn it into a cloud task.

A cloud job can still pause for missing information or permission. Labels and features vary by account and app version.

Describe the finished result, so the AI can do the work instead of only discussing it.

A useful request answers four questions: What do you want? What should it use? What may it change? What does finished look like?

Example · Customer interviews Turn messy notes into the next decision for your startup.

Read these customer interview notes and identify the three problems mentioned most often.

Use only the attached notes. Include one supporting quote for each problem, and flag anything uncertain.

Create a one-page summary with one small experiment I could run next week. Draft it here; do not contact anyone.

Make it a playbook: save a prompt that works, then reuse it with new inputs. You do not need a special feature to start a prompt playbook.

A Project holds the files and instructions an ongoing piece of work needs.

Use a Project for work you will return to. Create one in ChatGPT, add the relevant files, and start related chats inside it.

Example · Deep Space Field Notes Keep observing notes and equipment details together, so you can plan the next clear night without explaining your setup again.

Using the observing notes and equipment list in this project, suggest three targets for my next clear night. Ask for my location and date if they are missing. Explain why each target suits my equipment.

Add rules you should only have to say once

Put preferences in Project instructions. For a coding project, save them in AGENTS.md at the top of the project folder. That file gives Codex standing instructions.

Example · BullyBearAI Require evidence for claims about a financial app.

# Working rules
- Explain changes in plain language.
- Label sample data and uncertain claims.
- Never invent customer feedback or test results.
- Test changed behavior and report what was not checked.
- Finish with what changed, the evidence, and the next step.

A plugin or connector lets the AI use another app with your permission. Start with the apps that hold your work.

Connect one service first

1. Open the Plugins directory. Find the service, select + to add it, and sign in when prompted.
2. Read the permission screen. If it shows checkboxes, tick the access you want to grant, then approve.
3. In Settings → Plugins (called Apps in some versions), open the connection’s Permissions. Keep Always ask for actions you want to review. Start a new chat to use the plugin.

AWS Core adds AWS tools and guidance. Your AWS sign-in and account permissions determine what those tools can actually do.

Find recent customer questions in my connected Gmail and the product plan in Google Drive. Summarize the three biggest gaps between what customers ask for and what we plan to build. Link the sources. Only read; do not send or edit anything.

Why this matters: the AI can work from your actual business information. Reading a message and sending one are separate actions; approve the access and actions you intend.

A skill is a saved set of steps the AI can follow again. Use one for a task you repeat.

Example · Good Morning and Good Evening Start with a short plan. Finish with a record of what changed and what comes next.

Help me create two reusable skills: Good Morning and Good Evening.

Ask which calendar, email, chat, GitHub, and document connections to check, and where to save my daily note.

Good Morning: check those sources and give me three priorities, today's meetings, and blockers.

Good Evening: compare the morning plan with actual progress. Save completed work, open commitments, and tomorrow's first step.

Include source links and say when a connection could not be checked. Do not send messages or change my calendar. Test both skills once.

Set it up: choose @skill-creator in ChatGPT, where available, and paste the prompt. In Codex CLI or the IDE, use $skill-creator. Save the skills, then ask: “Run my Good Morning skill.”

A skill remembers how. A schedule decides when. Try the skill yourself before asking it to run automatically.

Sites turns a conversation into a working website, small app, or browser game that you can share with a link.

How to use it: choose @Sites and describe what you want. Ask for a first version, try it, then describe one improvement at a time. Choose the audience before publishing the version you want to share.

Example · Before the Morning Bell A scary escape room: explore an abandoned hospital, find a key, and unlock the exit.

@Sites Create a short, spooky browser escape game called Before the Morning Bell.

Start with one room, one hidden key, and one locked exit. Add a clear goal, readable controls for mouse and touch, and an ending when I escape. Include an optional sound toggle; keep sound off until I turn it on.

Save a first version for me to review before publishing. Include brief instructions for trying it.

Try the whole game. Then give specific feedback: “I found the key, but the door stayed locked. Fix that interaction and check the complete escape.”

A schedule runs a task later or repeats it. Browsing lets that task check current information on websites.

Example · Dumpster pricing watch Once a week, check rival companies’ public offers. Compare the same dumpster size, rental period, and included weight.

How to use it: give the AI the company websites and ask it to browse them once. Check that first result, then ask it to repeat the job at a specific time.

Check these dumpster rental companies in [city]: [official website URLs]. Compare [dumpster size], rental days, included weight, base price, and extra fees. Link each source and date the result. Mark unpublished prices “quote required.” Do not contact the companies.

Run this once now and save the comparison in [destination]. After I check it, schedule the same task for Mondays at 8 AM in [timezone]. Compare with the last successful result and report changes or failed checks. Confirm the task exists and where I can see its results.

Check the schedule in ChatGPT. A written promise is not a created task. Confirm its time, timezone, and first result. Use cloud-accessible files and tools if you want it to run with your laptop off.

Give the AI one outcome and enough direction to work through several steps. A goal describes the finish line; it is not a guarantee of unlimited runtime.

Example · BullyBearAI An overnight Codex session worked through epic #1090 and produced eight draft pull requests—proposed code changes ready for review.

The run reported 11h 25m of session elapsed time and a separate 2h 46m goal timer. It did not deploy to AWS. The Claude edition adapts the workflow; the original run used Codex.

1. Create an epic. This is the overall outcome, such as “Make our weekly customer report trustworthy.” Add small feature issues underneath, each with a clear completion check.
2. Label the features. AFK means “away from keyboard”: the AI has enough information and permission to finish. HITL means “human in the loop”: it needs a person’s decision, access, or real customer feedback.
3. Launch the ready work. Set the scope, time limit, and allowed actions. Subagents are extra AI workers; give each a separate task and have the lead agent track progress.
4. Review in a new chat. Give another agent the code changes and test evidence. Ask it to inspect and check the work before you use it.

Start in Codex with the project open and GitHub access available. Use /goal where supported, including the desktop app, interactive Codex CLI, or IDE extension. In Work on the web, paste the brief directly. Check Local or Cloud before stepping away.

In Codex, work through the AFK-ready features under [repository and epic]. Refresh the issues and project rules first; skip blocked or human-dependent work.

Use up to two subagents on independent tasks, with separate branches. For each feature, implement it, run the relevant tests, update documentation, and open a draft pull request. Save progress after each feature.

Stop after [time limit], at [usage limit], or when no ready work remains. Report any limit you cannot measure. Do not merge or deploy. Finish with PR links, test evidence, dependencies, and what needs a person.

Hand the result to a fresh reviewer

A second chat gets a clean look at the work. Include the epic, pull requests, and the original completion checks. If deployment is part of the review, name the exact test environment.

Review BullyBearAI epic #1090 and PRs #1128, #1129, #1130, #1134, #1135, #1136, #1139, and #1140. Fetch current changes and dependencies; do not rely only on the author's summary.

Check the code and rerun tests. Fix problems on the affected PR branch and push the updates. Recheck after every fix.

Deploy and test sequentially only in [authorized AWS test account, region, and environment]. Stop if that target is unspecified. Record the deployed commit and results for each PR. Do not merge or touch production, broker access, or trading controls. Finish with the review order and remaining blockers.

Before you leave: check that the agent can reach the files and tools, knows when to stop, and has somewhere to save progress. Long jobs can still stop for limits, questions, or missing access.