Unit 02.01: Instructions that constrain rather than encourage

Half the instructions in a typical agent prompt change nothing about what it does.

Name the output each line makes impossible

An instruction earns its place if you can point at a behaviour it rules out. Encouragement rules out nothing.

The table below sorts six instructions.

instruction                                            rules something out?
Be helpful and thorough.                               no -- encouragement
Use only the invoice record and the contract.          yes
Try to be accurate.                                    no -- encouragement
If either document is missing, stop and say which.     yes
Never estimate a figure that is not written down.      yes
Do your best.                                          no -- encouragement

3 of 6 would change what the agent does

"Be helpful and thorough" and "try to be accurate" cost tokens on every call and rule out no output at all. "Never estimate a figure that is not written down" rules out an entire class, and - usefully - you can check afterwards whether it held.

That checkability is the second benefit. A constraint you can verify becomes a test; an encouragement can only be hoped for.

The mistake this prevents

The mistake is adding instructions in response to a bad output without removing the ones that did not work. Prompts accumulate - three rounds of "please be careful about X" and the genuine constraints are buried among sentences that change nothing.

Takeaway

Every instruction should name an output it makes impossible. If you cannot name one, delete the line - it is costing tokens and diluting the instructions that work.

Mark complete