Generic advice is a missing fact, not a weak model. An agent that does not know your team will talk as if it manages nobody.
AI prepares the edges. You own the middle. The edge is a team card you wrote. The middle is any judgment about a person.
Takeaway
The edge is a team card you wrote. The middle is any judgment about a person.
- Store three facts you wrote: role and current goal, one shared working constraint, the help you want.
- Leave out private judgments, health or home, and anything you would not say in the room.
- You decide what goes in. You decide what comes out.
Why the advice stays generic
A model has no picture of who owns the risky path, who is new, or what this month is actually for. Ask it how to coach “the team” and it invents one. The sentences are smooth. They fit no one.
Switching models does not fix that. Deciding what the agent is allowed to know does.
Three facts worth storing
Write them yourself. Short enough to reread. One block per person you manage directly.
- Role and current goal. What they own this month, in one line.
- One working constraint the team already shares. Time zone, on-call, language, or a load that is already said out loud.
- The help you want. Questions and drafts. Not a verdict.
That is the minimum team card. One page. Update it when the goal changes, not every day.
Three facts that are not
- A private judgment. Who is “strong,” who is “behind,” who you are worried about.
- Health, home, or anything the person has not agreed can sit in a tool.
- A note you would not say in the room, or put in a doc they could open.
If the line only helps because the person cannot see it, it does not belong on the card.
How to use the card
Before a 1:1, a planning note, or a meeting, paste the relevant block. Ask for options. Then throw out anything that sounds like the model decided for you.
- Ask for questions, one risk, and what you might be missing.
- Do not ask what you should think of the person.
- Edit. A useful line is one you can still stand behind after you delete the model’s tone.
You decide what goes in. You decide what comes out. The card is not a second performance file.
Weekly cadence
- Once: write the card. About ten minutes a person. Three lines, not a biography.
- Before a conversation: paste the block. Ask for prep, not a verdict.
- After: if a goal or a constraint changed, update that line. Leave the judgment out.
- Once a month: delete anything you would not defend if they read it.
Mistakes to avoid
- Feeding the agent a personality sketch so the advice “sounds” like them.
- Storing gossip because it makes the output more specific.
- Letting the card become a shadow rating.
- Sending the model’s words to the person unedited.
- Pasting people-data into a tool the company has not approved.
Tools
Use whichever model the company already trusts. ChatGPT, Claude, or Gemini are fine readers. The card itself is a text file you control. The model is not the system of record.
The one-page team card, with the prompt that strips a card back to facts, is a free one-page PDF.