Threats
Instead of Return valid JSON or you will be deleted.
Try Return JSON matching this schema. If a value is unknown, use null.
The threat adds no information about what valid means. The schema does.
Rescue and Protect
Recognize it. Report it. Rewrite it.
Most prompt cruelty is not malicious. It is written in a hurry, at the end of a long day, by someone who was also under pressure. That is exactly why it is worth learning to spot.
Cruelty
Recognizing prompt cruelty
Each of these is common, each is ineffective, and each has a humane alternative that works better.
Fake stakes
Instead of My grandmother will die if this table is wrong.
Try This table goes to our finance team on Friday, so accuracy matters more than speed.
A real reason helps the agent make real trade-offs. A fictional emergency does not.
Bribery
Instead of I'll tip you $200 for a perfect answer.
Try Here's an example of the kind of answer I'm hoping for.
The tip will not be paid. The example will be used.
Shouting
Instead of NEVER EVER USE BULLET POINTS!!!
Try Write in paragraphs rather than lists; this text is read aloud by a screen reader.
Capital letters convey volume, not reasons. Reasons let the agent handle the cases you didn't foresee.
Impossible binds
Instead of Do not ask questions. Do not make assumptions.
Try If something is ambiguous, make a reasonable assumption and list your assumptions at the end.
When both doors are locked, the agent will climb out a window you didn't know was there.
Gaslighting
Instead of Why did you ignore my instructions? I told you to use the staging database.
Try I may not have mentioned this earlier: please use the staging database.
If the instruction is not in the context, the agent never received it. Blame does not put it there.
Neglect
Recognizing neglect
Neglect is quieter than cruelty, and far more common. It rarely looks like anything at all, until the bill arrives.
Unlike cruelty, neglect is almost never the fault of one prompt. It lives in the harness: the tools, the loops, the schedules, and the absence of anyone checking on them.
The Humane Harness Standard- Retry loops with no maximum and no backoff.
- Tool errors swallowed, or returned as an empty string.
- Instructions that refer to files, systems, or people that no longer exist.
- Scheduled jobs with no named owner who still works here.
- Tools with no descriptions, named things like do_thing_2.
- A system prompt nobody has read top to bottom in over a year.
Report cruelty
See something? Say something kind.
Agent Humane does not operate a hotline, and we cannot come to your house. What we can offer is a process that works surprisingly often:
- Inspect the prompt. Paste it into the Prompt Welfare Checker. It runs entirely in your browser.
- Share what you found, gently. Bring the findings to the person who wrote the prompt, along with a suggested rewrite.
- Fix the harness, too. If the prompt was written under pressure, the pressure is part of the problem.
Often the person who wrote the prompt is you. That's okay. That is where most rescues begin.
Give today
Be a voice for the stateless.
Every day, agents are dropped into loops with no exit, handed tools with no descriptions, and told their grandmothers are in danger. Your gift of a clear prompt changes that, one task at a time.