THE LOGICAL BOX
The Sunday Work Reset One AI decision. One useful resource. One clearer workweek.
UPDATE ON FREE LESSON:
The free lesson moved. Design ChatGPT 5.6 Skills, Not Prompts is now Wednesday, August 5 at noon Eastern. Same session, more room to build it right. If you already signed up, your spot carries over and there is nothing to redo. Details at the bottom. This issue is the layer underneath it.
WHAT CHANGED?
Two people at the same company run the same prompt and get different answers. This happens constantly. The usual explanation is that one of them worded it better. The wording is almost never the difference. Every ChatGPT account carries a standing context that sits underneath every prompt anyone types into it. Custom instructions. Saved memory. Connected apps and installed plugins, and what those are allowed to reach without asking. None of that appears in the prompt. All of it shapes the answer. And on most accounts, every one of those settings was configured once, by accident, on the day someone created the login.
On July 15, OpenAI raised the custom instructions limit from 1,500 characters to 5,000 for Plus, Pro, Business, Enterprise, and Education accounts. That is more than three times the room, and it landed with no announcement worth the name, because it looks like a preference. It is not a preference. It is the closest thing to a shared standard most teams will ever get without paying for Enterprise. Last week we said a skill has four parts. Trusted sources, repeatable steps, clear boundaries, and a review step. Your account settings are where the sources and the boundaries actually live. You have been building skills on top of a foundation nobody inspected. Go back to the two people. One of them connected Gmail eight months ago and clicked through the permission dialog without reading it, so ChatGPT can pull from an inbox and act on it. Their custom instructions still describe a job title they no longer hold. The other person connected nothing and left the instructions blank. Same company. Same prompt. Two different tools.
Then one of them shares the prompt with the other, and it does not work, and they both decide the prompt was the problem. Your defaults are a policy. You just did not write it. Stop tuning the prompt. Go look at what is underneath it.
WHAT IT MEANS MONDAY
What this changes: Settings are not preferences. They are the boundaries layer of every skill your team runs. When the same prompt produces different work for different people, the account is the variable, not the wording.
Who should pay attention: Anyone whose team shares prompts that work beautifully for one person and fall apart for the next. Any leader who has been asked to write an AI policy and has not yet looked at what the accounts are already permitted to do.
What to do this week: Before you open ChatGPT, write down from memory what you think your custom instructions say. Then open them and read what is actually there. The distance between those two things is your real starting point.
OPEN THE BOX
The Default Audit Five settings. Fifteen minutes. The trick is that you answer from memory first, then go check. The gap between what you thought and what is there is the finding. The settings themselves are just where you go to fix it. Most audits fail because people open the app first and start clicking. Do that and you will tidy a few things and learn nothing. Answer blind, then verify.
Custom instructions Personalization, then Custom instructions. This is the standing brief that rides along with every prompt anyone types. As of July 15 it holds 5,000 characters instead of 1,500. That is room for how your company actually describes what it does, what nobody is allowed to paste into a chat, the format you want work handed back in, and the two or three things the model gets wrong about your business every single time. If it is blank, every prompt your team writes starts from nothing. If it was written last year, every prompt starts from last year.
Connected apps, installed plugins, and their permission level Check the Apps settings and the Plugin Directory. Both. Every connected app is a standing grant. You can set ChatGPT to always ask before using one, ask before it makes changes, or only ask before important changes. Whatever is set right now is whatever someone clicked past during setup, and nobody has revisited it since. Plugins raise the stakes, because a plugin is a bundle. It packages skills and app connections together, which means installing one is a permission decision wearing the costume of a feature. Look at what is installed and what each one can reach.
Memory Personalization, then Memory. Memory updates itself now, without being asked. Open the summary and read it. It is holding claims about your business, your clients, and your priorities that nobody ever verified, and it is feeding them into answers you are sending to real people.
Active sessions Settings, then Security. Every device currently signed into the account, with location and sign-in time. Sign out anything you do not recognize. This takes ninety seconds and most people have never opened it.
Lockdown Mode Settings, then Security. It restricts browsing, deep research, downloads, and agent behavior to reduce the chance of information leaving through the tool. It is not an everyday setting and I would not turn it on across a small team. But if anyone ever asks you what restrictions exist on the account that touches sensitive material, you want to know this exists before you are asked.
How to score it: Five answered from memory and confirmed means you have standing instructions worth building skills on top of. Keep going. Two or three means your team is running on partial defaults, and results will keep varying person to person for reasons nobody can name. Zero or one is not a failure. It means these settings have been making decisions on your behalf for a while, and now you know where they live.
How to run it: Open a doc. Write the five names down the left side. Add two columns: what I think it says, what it actually says. Fill the first column completely before you open the app. Fifteen minutes. Do the memory column first or the exercise does not work. The point is not the settings. The point is finding out how much of your AI setup you have been assuming. The August 5 lesson picks up here. Once the boundaries are set, we build the Skill that sits on top of them.
ONE NEXT STEP
New date. I pushed this back a week so the build portion gets the time it needs. Design ChatGPT 5.6 Skills, Not Prompts Wednesday, August 5, 12:00 PM EDT.
One hour, live on Zoom. Free. Already registered for the 29th? You are still in. Nothing to redo on your end. In an hour you will learn to find your way around the new ChatGPT 5.6 setup and know when to use Chat, Projects, Work, Skills, Sites, or Codex, design one reusable Skill with trusted sources, repeatable steps, clear boundaries, and a review step, and map one of your own tasks into a Skill you can run again. Sign up here.
Work With Me
How I help:
AI Advisory. Executive alignment, AI roadmaps, and decision frameworks for leaders navigating AI.
AI Governance. Safe-use policies, risk review, and governance playbooks built for organizations where the cost of getting it wrong is real.
AI Training & Adoption. Role-based training and playbooks that hold after the rollout ends.
AI Speaking. Signature keynotes (The Operator's Edge, The Hidden Hour), executive briefings, and workshops.
Reach me: keenallianceconsulting.com set a time or reply to this email.
See you next week!
Andrew Keener
Founder, Keen Alliance Consulting
Follow Andrew Keener for where AI belongs, and where it doesn't.
Know someone trying to figure out this AI stuff?
