Brief ChatGPT Agents With Hard Limits

When ChatGPT agent features are available on your plan, you can give the tool more freedom than a regular chat. An agent can browse, run code, or call services on your behalf. That freedom is useful, but it is also where things go sideways if the brief is too loose. A good brief sets hard limits before the agent starts.

Hard limits are not suggestions. They are walls the agent cannot cross. You name the goal, you name the tools it may use, you name the actions that require your approval, and you name the actions it must never take. When the brief is tight, the agent works fast and stays safe. When the brief is loose, the agent improvises, and improvisation is where surprises live.

What a hard limit is

A hard limit is a rule the agent must follow no matter what. It is different from a preference. A preference says try to keep it short. A hard limit says do not write more than 100 words. Preferences bend under pressure. Hard limits do not.

When to set hard limits on agents

• The agent has access to external tools or services
• The agent can make changes that are hard to undo, such as sending a message or deleting data
• You are running the agent for the first time on a new task
• The task involves sensitive data such as financials or personal information
• Other people will see the output without your review

Prompt for an agent with tool restrictions

Goal: research the top 5 competitors in the meal-kit delivery space and return a comparison table.
Allowed tools: web browsing only.
Hard limits:
- Do not sign up for any service.
- Do not enter payment information anywhere.
- Do not download files from unknown sources.
- Do not visit more than 10 pages total.
Approval needed: none, this is read-only research.
Output: a table with columns: Company, Price range, Delivery area, Unique feature.

Why this prompt works

The goal is clear and scoped. The allowed tools section limits the agent to browsing. The hard limits prevent side effects like signups or downloads. The approval line says none is needed because the task is read only. Every boundary is stated before the agent moves.

Prompt for an agent with approval gates

Goal: draft a response to each unread support ticket in my inbox.
Allowed tools: read email, draft replies.
Hard limits:
- Do not send any email. Only save drafts.
- Do not access emails older than 7 days.
- Do not include discount offers or refund promises.
Approval needed:
- Before saving any draft, show me the draft and wait for my "approved" reply.
Output: for each ticket, show the original subject, your draft reply, and a confidence note.

Prompt for an agent that must not take certain actions

Goal: organize the files in my project folder by type.
Allowed tools: file reading and moving.
Hard limits:
- Do not delete any files.
- Do not rename files, only move them.
- Do not create new folders beyond: documents, images, code, other.
- Do not touch any file in the .config directory.
Approval needed:
- Before moving more than 10 files, pause and show me the list.
Output: a summary table showing how many files moved to each folder.

What belongs in a good agent brief

• A one sentence goal that says what done looks like
• A list of tools the agent may use and nothing else
• Hard limits on actions that could cause harm or surprise
• Approval gates for any action that is hard to reverse
• The format for the final output so you can review it quickly

How to refine the limits over time

After the first run, review what the agent did and what it skipped. If it stayed well within the limits, you might loosen one. If it bumped against a limit in a way that felt unsafe, tighten it. Good limits evolve with experience, but they should always start strict and get looser, never the other way around.

Common mistakes

• Giving the agent a goal without any limits, which lets it improvise freely
• Using soft language like prefer not to instead of do not
• Forgetting to set approval gates for actions that affect real data
• Listing too many tools as allowed when the task only needs one
• Trusting the agent on the first run without reviewing its actions

How to check the results

Read the action log if one is available. Confirm that the agent stayed within the allowed tools and did not trigger any hard limits. Then review the output against the goal. If the output matches and the log is clean, the brief worked. If either is off, tighten the brief and run again.

Takeaway

An agent is only as safe as its brief. Start strict, set approval gates, and loosen the limits only after you have seen enough clean runs to trust the pattern.