--- name: never-hand-off-checks description: 15 rules from the Noesa course "What to never hand off". For professionals delegating daily work to AI who need a clear, defensible line: what AI drafts, what needs their judgment and sign-off, and what never goes to it. --- # What to never hand off — the rules Use with: Claude Code or Claude (save as a skill), Cursor (save under .cursor/rules as .mdc), ChatGPT or any other assistant (paste the text below into custom instructions or a project's instructions). 15 rules, taken from the course at https://noesa.leafsoft.online/c/never-hand-off Each heading is one thing the course teaches. Most are checks to run on your own output before presenting it as done; a few are background you are expected to have. "Wrong by default" lists 6 specific ones. Apply these to the thing you are producing — the type, the schema, the query, the copy — not only to how you explain it. Where a rule names a field, a format or an identifier, that name belongs in the output. ## Meet the intern with no stakes Explain why AI can perform work without owning its consequences and classify one task by who remains accountable. ## Separate the draft from the decision Mark where an AI-produced draft ends and a human decision begins in a workplace task. ## Keep the facts you sign for Identify factual claims inside AI-assisted work and state what your sign-off promises about them. ## Keep the judgment calls Distinguish a useful AI comparison from a judgment call that requires human context, values, and authority. ## Keep the relationships Identify relationship-bearing communication and choose an AI role that preserves human presence and trust. ## Keep the confidential Classify an AI task separately from its required material and set a delegation boundary that changes when the inputs change. ## Watch the stakes, not the task Classify the same task differently by consequence, reversibility, reach, and vulnerability. **Wrong by default:** - Reach is one of the four things that set stakes, and a template multiplies it rather than reducing it: one wrong date in a message that goes to thousands of customers is that many wrong dates, sent at once and unrecallable. Judge the consequence at the volume the thing actually ships to. - It is the message a stranger imitates to take an account, and it reaches whoever is currently locked out. Vulnerability of the reader and the security weight of the message set the stakes here, not how repetitive the wording is. - Reversible means the harm can be undone, not that another message can be sent. A correction reaches fewer people than the original, arrives after some have already cancelled or budgeted, and a price commitment in writing may bind you. ## Mind the compliance lines Recognize legal, financial, health, safety, privacy, and compliance claims that require an authorized or qualified human owner. ## Spot automation creep Identify when one-off AI assistance has become an automated decision or communication channel without adequate review. **Wrong by default:** - Do the arithmetic: 5% of 200 is 10 reviewed and 190 sent to customers unread every week, about 2,470 a quarter. A sample looks at what already went out, so it is an audit after the fact, not a checkpoint before a consequential message. - A model's confidence is about its own wording, not about what happens to the customer if it is wrong. It has no view of the stakes, so it cannot be the thing that decides whether a human sees the message. Route on consequence instead. - Stakes accumulate through reach. A fault in the drafting step is not one bad reply, it is the same bad reply 190 times a week, and nobody read any of them. ## Stay the author of your thinking Use AI for expression and challenge without outsourcing the initial reasoning your judgment depends on. ## Design the human checkpoint Place a named, meaningful human checkpoint at the moment an AI-assisted workflow becomes consequential. ## Write the escalation rules Write escalation triggers based on severity, uncertainty, and novelty and route each trigger to a named owner. ## Own the "AI was wrong" day Respond to a shipped AI-assisted mistake by owning the impact, correcting it, and strengthening the workflow without hiding behind the tool. ## Draw your own red lines Assemble a defensible Hand-off Charter that classifies tasks and specifies conditions, owners, checkpoints, and escalation routes. ## Decide a week of real work Classify a real week of tasks with your Hand-off Charter and defend two decisions using evidence, stakes, and ownership.