Lesson 5.85.8 The same brief on today's tools

In everyday life. The same shopping list works in two supermarkets, as long as it says "the small bag" rather than "the usual."

At work. The same payment-run brief runs in Claude, then in ChatGPT, with the same outcome, rules and checks. Both put the run file where the brief says.

The four parts describe the job, not the tool, so one brief should run on any capable worker. The three boxes below follow the same eight lines in the same order, and the open-source box adds a ninth, on running it safely. Read the box for the app you use. The others are there to compare. A number such as (5.3) points to the concept that line matches.

Anthropic, as verified 6 October 2026, updated 31 October 2026.

• Where you brief it. Any conversation in the new Claude experience, with no mode to pick. As of 31 October 2026, it is being released first to Pro, Max and Ultra plans. Pro, Max and Ultra accounts that do not have it yet still see Chat and Cowork.¹
• What to include (5.1). Claude's help center lists the desired outcome, the format, and the inputs Claude will need.² Its guide to Docs adds what the result is for, who will read it and what it should cover. It also says to name the file or app where the content is kept.³
• Outcome or steps (5.2). Describe the outcome, and Claude works through the steps. You do not need to break the work into steps.²
• Context (5.3). Attach files and links, or name connected apps such as Google Drive. Google Docs added to a chat or project sync from Drive, so Claude reads the latest version.⁴
• Plan first (5.4). When you ask for a doc, Claude asks clarifying questions before it writes a draft.³ Manual mode also asks before each action.²
• Helpers (5.4). For a complex task, Claude may on its own coordinate several sub-agents working at the same time.⁵ You can ask for helpers in plain words, or ask Claude not to use them.⁶ In Claude Code, a helper starts fresh and does not see your conversation, unless it is a fork, which starts with a copy of it.⁷
• Deliverables and destination (5.7). Documents, spreadsheets with working formulas and presentations.² Claude Docs is on every plan, Free included, and you can export a doc to Word, PDF, Markdown or Google Docs.³ Anthropic's artifacts page says Slides and Design are on every plan too.⁸ Its page on the one conversation still calls all three beta, on paid plans.⁹ With the Google Docs, Sheets and Slides connectors, Claude creates and edits Google files. Name the destination you want: a Google file, a Claude Doc or a file to download.⁴
• Autonomy. A permission setting in the message box. Manual asks before each action. Auto keeps working, with automated safety checks before actions. It covers the whole conversation, and you can stop or redirect Claude.² As of 3 November 2026, Auto is the default on Pro, Max, Ultra and Team, and Manual is the default on Enterprise.⁹ Anthropic's launch post says Claude asks by default, so read what your own message box shows.¹⁰

OpenAI, as verified 6 October 2026, updated 4 November 2026.

• Where you brief it. A ChatGPT conversation. If your app still shows Chat and Work, choose Work.¹¹
• What to include (5.1). OpenAI's prompting guide lists goal, context, output and boundaries, using only the parts that help. For a longer task it adds: the result, the source material, the audience, and how you will review the work.¹²
• Outcome or steps (5.2). Start with the result, not a detailed list of steps. Describe a process only when the process matters.¹²
• Context (5.3). Attach files, use a project, or name connected sources and what to take from each. Add only the sources that matter.¹²
• Plan first (5.4). Ask for a plan when the approach matters.¹²
• Helpers (5.4). ChatGPT and Codex can run subagents. Usually, you ask for them directly, such as "check these in parallel". In Codex, a project's or skill's instructions can also ask for them. The desktop app shows each helper's thread, with its work and the summary it sent back. Helpers keep the permission setting you chose, and they use more of your plan than one worker doing the same job.¹³
• Deliverables and destination (5.7). Documents, spreadsheets and presentations, including native Google Docs, Sheets and Slides when the Google Workspace app is turned on. OpenAI says to name the output format and where the file should be created.¹⁴ On Pro, Business and Enterprise, ChatGPT also makes pages in Space: editable documents you can share and edit together. Name the space when you ask for a page.¹⁵
• Autonomy. Boundaries you write in the prompt, such as requiring your approval before ChatGPT sends, publishes or changes information others rely on.¹² In the desktop app, a permissions control sets local actions. Ask for approval is the recommended start. Two other choices, Approve for me and Full access, must be turned on in settings first. In settings, Approve for me is called Auto-review.¹⁶

Open source, as verified 23 October 2026.

• Where you brief it. OpenCode, an open-source coding agent, in its terminal, desktop or web app.¹⁷ All three connect to one OpenCode service on your own computer.¹⁸
• What to include (5.1). The same four parts. Rules for every task, such as "always cite the policy," can go in an AGENTS.md file in the project's folder. OpenCode loads it for you.¹⁹
• Outcome or steps (5.2). Describe the outcome. Build, the default agent, does the work.²⁰
• Context (5.3). The files in its folder, called its workspace, and other apps, through MCP servers, which work like connectors.²¹ Name the governed source by its file name.
• Plan first (5.4). Choose the plan agent. It plans without editing your project files, but its shell commands still follow your rules.²⁰ Approve the plan, then switch to build.
• Helpers (5.4). Built-in general and explore helpers. Each runs in a child session with a fresh context, not the parent's conversation, and with its own permission rules.²⁰
• Deliverables and destination (5.7). Files it writes on your own computer, such as the proposal CSV.¹⁸ Name the folder and the file name.
• Autonomy. Every agent starts out allowed to use every tool. Only reading a file of sensitive settings asks first. A rule can set each action, such as editing files or running commands, to allow, ask or deny.²² A custom agent, one text file, carries its own instructions, model and rules.²⁰
• Running it safely. Before the first brief, set every action that can change or send something to ask, so nothing happens until you approve. Helpers have their own rules, so set theirs too.

The comparison. The AI vendors' guides use other words for the same parts, and they put some instructions in other places. Anthropic's course for the Claude Certified Associate exam lists five parts.²³ OpenAI's prompting guide lists four.¹²

The book's parts match Anthropic's course on length and tone, which OpenAI's guide puts with the output. OpenAI's boundaries cover both halves of the book's constraints. Anthropic's course has no part for what a worker may do alone. The course's role belongs in the worker's definition, which both AI vendors call a system prompt (5.1). Each product also has a permission control for autonomy. Both have their own editable documents (Claude Docs and ChatGPT pages), and both can create Google files. On both, the destination you name decides where a file is created.

One difference changes the brief. Both let you keep a person involved, but they set the limit in different places, and their defaults differ. Claude's Auto, the default on most paid plans, keeps working with safety checks. Its Manual asks before each action. ChatGPT's desktop Ask for approval can read and edit files in its workspace, the folder it works in, without asking. It asks only before going beyond that. So state in words which actions must wait for you, such as "Change nothing and send nothing." The other differences are about access: which plans include it, and what is still in beta. OpenAI's page says Work, its name for longer tasks, needs an eligible paid plan,¹¹ and its pricing page lists some on the free desktop app.²⁴

Open source widens that difference. OpenCode's default agent asks first only before reading a file of sensitive settings, and it may work outside its folder.²⁰ So write the autonomy half of the constraints in words, and set matching rules in the tool before the brief runs.

Helpers differ too (5.4). Claude may start helpers on its own when a task is complex. ChatGPT starts them when you ask, and Codex also when a project's or skill's instructions call for them. So on ChatGPT, a brief that wants parallel work must ask for it, and on Claude, a brief that does not want it must say so. On all three, write every rule a helper needs in the brief itself.

Steering a run (5.5). Claude Code holds a message you type while it works, and passes it on as soon as its current tool calls finish, in the same turn.²⁵ In Claude Tag, anyone in the channel can steer a running session by replying in its thread.²⁶ In both, name the part that changed.

When the model changes (5.5). Run each saved brief again, on either AI vendor or an open-source tool. Anthropic's guide for Claude Opus 5 says the model checks its own work without being told to. Old lines that ask it to verify make it check twice, so the guide says to remove them.²⁷ Anthropic's guide for Claude Opus 5.5 adds that lines telling the model to think carefully can go too, because how hard it thinks is a setting (5.1).²⁸

Figure 5.5. The same brief on both AI vendors, as verified 6 October 2026 and updated 4 November 2026. The two gold rows are the ones to state in words on both AI vendors: where the file goes, and how far the worker may go alone.

What stays the same. The outcome, the rules and the checks you will run stay the same on either AI vendor or an open-source tool. Delivery and access details may change: where a file is created, how inputs attach, how permissions are set. If you have to change the outcome or the rules to move a brief, it was describing a tool, not a job.