Handing a new hire a URL and saying "ask the bot" creates shadow IT habits. A one-week plan sets norms for prompts, privacy, and when to escalate to humans.
Day-by-day sketch
Day 1: account access, acceptable use, and where logs live. Day 2: guided examples for the top three job tasks. Day 3: practice with citation-required answers on internal docs. Day 4: failure modes—refusals, hallucinations, and how to file a bad answer. Day 5: shadow a teammate reviewing bot output. Day 6–7: independent work with a short retrospective note.
Guardrails to teach early
Never paste secrets. Prefer project folders that are already approved for RAG. Do not treat local as "unsaved forever" if logs exist. Know who owns the prompt and the model version.
Resources
Point people at tutorials/basics, product surfaces like MicroAI One, and policy-ish methodology at methodology. Hardware curiosity can wait, but link match for the curious. Private deployment context: solutions.
Extra curriculum
Include a five-question quiz graded on process, not eloquence. Store onboarding prompts in the same versioned library as production. Invite security to a fifteen-minute slot on phishing via pasted text. Revisit the plan each quarter as tools change. If you need a structured service wrapper, see services or contact.
Onboarding is cheaper than cleaning up a month of bad norms.
Manager checklist
Confirm the new hire completed the quiz before granting production RAG groups. Pair them with a buddy for the first bad-answer report. Collect anonymous feedback on whether the bot norms feel realistic. Update the week plan when the default model changes. Keep accessibility in mind: keyboard-only paths and readable contrast in the UI you standardize on.
Content of the quiz
Ask where to report a bad answer, what never to paste, how to read a citation, and which model is approved for customer email drafts. Fail the quiz on any secrets question. Offer a retake after reading the one-pager. Managers should not waive the quiz for senior hires—senior people paste secrets too. Keep quiz questions versioned with the onboarding plan. Add a short accessibility note so screen-reader users know the supported UI path. Revisit examples when your default model or RAG corpus changes so training does not describe a ghost product.
Published by MicroAI One (本地AI 1号网), a Shenzhen team working on local LLM hardware, software and deployment.