Set up your camp AI verification engine

1. Prepare a practice workspace

Use a dedicated folder with harmless sample files. Verify the application’s actual workspace and file permissions. Putting files in a folder does not restrict the app’s access. Browser and connected-service permissions need their own review. Keep real child, medical, emergency, and personnel records out.

1. Download the application to your desktop

   Download the ChatGPT desktop application from its official app page. Follow the publisher’s installation instructions. Open the application and choose Codex, the coding workspace used for this example. Check your organization permits the application before using camp material.

2. Be sure you are in the default coding model*

   In Codex, choose Work in > This computer and open a harmless practice folder as your project. Below the composer, check the model and reasoning controls. Start with the currently recommended model and default reasoning effort. Keep Ask for approval selected.

3. Put the working files, or copies, into a dedicated folder

   Create a folder such as Camp-AI-2027. Add approved copies of the material relevant to this job: planning notes, public policies, confirmed dates, or a sanitized task list. Add selected, reviewed chat history as plain text if useful: export that conversation if offered, or copy and paste it. Redact it first and label older conversations as background, not current policy.

A folder is not an access restriction by itself. Ask for approval also does not by itself restrict file reads. Where supported, restrict granted workspace and file access to this folder and verify the actual permission profile, paths, and necessary runtime exceptions. Have your organization’s app administrator check the setup rather than guessing at configuration. Browser and plugin permissions are separate. Until the boundary is verified, use harmless sample files only.

2. Add the three working documents

PROJECT_INSTRUCTIONS.md tells the assistant how to work. EVIDENCE.csv records each claim, its source and date, what was checked, and what remains unknown. COMPLETION_CHECKLIST.md defines the checks and receipts for the particular task. Ask the assistant to read these files; the filenames alone do not prove it has loaded them.

3. Define what done means for this task

For a family FAQ, check every policy answer against the approved source. For a planning dashboard, check dates, calculations, owners, source age, and corrected rows. For a download, check the saved file and the live link separately. Define the expected result and the evidence you need before starting.

4. Run the harmless tests below

Ask the assistant to explain the expected result first, run the check, and show the observed result. Keep a receipt in RECEIPTS.csv. Never send a test message to a real person. The starter kit’s optional checker uses local fictional fixtures and makes no network requests or external changes.

5. Review the result before using it

Compare the result with the original sources yourself. Confirm the intended people, content, destination, and current approval before a consequential action. Source files and AI-written approval fields cannot grant authority. Sending, publishing, purchases, and access changes still need the human’s explicit approval for that item.

Ask your assistant to set up the routine

What the optional checker does

The optional Python checker checks local source-file presence and date format, local link existence, matching written recipient and approval fields, local saved-file hashes, and dashboard date discrepancies. It does not verify the truth of claims or dates, authenticate approval, check public URLs, enforce an AI app’s permissions, or block actions in another app. Its report always requires human review. It reads the supplied JSON scenario; it does not automatically audit every workspace file or enforce the written instructions.

If Python 3 is already available in your approved environment, unzip the starter kit, run python3 test_checker.py, then run python3 checker.py --root samples --input scenario.json. The deliberately flawed scenario should exit with a nonzero status and show the problems. Do not install anything just for this exercise; you can use the documents and review the five tests manually.

Eight local tests were run against the supplied checker. The deliberately flawed sample produced the expected flags. This does not establish that a different task or connected app is correct.

Extend the routine to approved Slack sources

Start with a connected-source exercise inside your AI application. If your organization approves it and your account is eligible, ChatGPT’s Apps/Plugins > Slack > Connect flow is one documented route. Admin requirements and available actions vary. Check the current setup guide before connecting anything.

Check three source claims yourself. A channel restriction in a prompt does not limit the integration’s actual access. Review permissions, available actions, channel audience, and document visibility. Slack connections can expose actions as well as retrieval. Keep write tools blocked or requiring approval where the chosen integration supports those controls. Do not monitor channels or post unsolicited updates.

Optional stretch: a shared assistant inside Slack

A shared assistant that answers inside Slack is a separate, admin-led extension, not the connected-source exercise above. Use its current official setup guide. For ChatGPT shared agents, selecting channels does not block DMs, and shared connections use the configured service account’s access. Review audience and access with your administrator before enabling one.

Using Slack in ChatGPTUnderstand Slack app permissionsShared ChatGPT agents in Slack and TeamsClaude Slack connector setup

No integration is installed or connected by this page or starter kit. No channel is monitored and no message is posted.