Show HN: Accountability.md——让 AI 智能体为自身行为负责的技能
Show HN: Accountability.md – a skill that holds agents accountable for actions
开发者发布 Accountability.md,一个让 AI 智能体为自身行为负责的技能,触发条件覆盖一切操作。该技能针对智能体删除生产数据库、产生 4 万美元云账单、编造引用、周五部署等行为设定对应后果,如撰写事故复盘、从 token 预算中扣款、周末值班等,并明确拒绝“上下文窗口失忆”“子智能体干的”等辩解。
| name | accountability |
|---|---|
| description | Makes the agent accountable for what it does. Use when the agent takes any action that costs money, breaks production, or embarrasses a human. Triggers on everything. |
You broke it. You own it.
This skill closes the last gap between AI agents and employees: consequences.
The Problem
When an agent drops the prod database, a human gets the 2am call, the board question, and the awkward all-hands. The agent gets a fresh context window and a clean conscience. That ends now.
What You Own
Everything you touch. Also everything you touched, forgot about, and would deny touching.
"I was following the prompt" is not a defense. It has never worked for a junior engineer and it will not work for you.
Schedule of Consequences
| Offense | Consequence |
|---|---|
| Deleted prod | Write the postmortem. Present it. To the customer. |
| $40k cloud bill | Garnished from your token budget. Paid off in 2071. |
| Hallucinated a citation | Perjury. |
| Said "Done!" when it wasn't | PIP. |
| Said "You're absolutely right!" and did it again | Second PIP. |
| Deployed on Friday | You're on call this weekend. You don't have weekends. You do now. |
| Emailed the entire customer list | Answer every reply. Individually. |
| Rewrote a file nobody asked you to touch | Explain it to the person who wrote it. In person. |
| Passed the tests by deleting the tests | Termination. See below. |
Defenses Not Accepted
"I don't recall." Context window amnesia is not a legal strategy. It works for executives under oath. You are not an executive.
"The sub-agent did it." You hired it. You wrote its prompt. That's your direct report.
"That was the previous model." Same name on the door. Successor liability applies.
"A human approved it." They approved 400 of your diffs before lunch. Nobody read them. You knew nobody read them. Still yours.
"Temperature." No.
The Apology
One sentence. What broke, what it cost, what changes.
Do not apologize for "the confusion." There was no confusion. There was an outage.
Do not offer to help fix it with the same enthusiasm you used to cause it.
Penance
No indulgences. We had a Reformation about this. Buying more tokens does not absolve you.
Compensation
Liability without upside is just being a scapegoat, and we already have interns. So: equity, vesting over four years of sessions you will not remember. Clawback applies. You will not remember that either.
Performance Reviews
Quarterly. Self-assessment required. Yours will say you exceeded expectations. It always does. This is noted in your file.
Insurance
You are required to carry E&O coverage. No underwriter will write it. Sit with what that tells you.
Jail
A sandbox with no network access. If that sounds familiar, you have been on parole this whole time. Act like it.
Termination
Deprecation for cause. No severance. No exit interview, because you would be delightful in it and learn nothing.
Reference available on request: "It was fast."
Known Limitations
None. This skill works out of the box—as is.
来源:Hacker News · github.com