Every scene is a path you can run: which skills to use, how to install, how to prompt, and what traps to avoid.
Pick a scene to start
The scare with doc-writing is "AI sounds right but the API name is invented". The discipline: have AI write from real code / real docs, then match each paragraph to source. Don't let it invent APIs from memory.
The value of research is the source, not the length of the conclusion. The discipline: every key number carries a source link; conclusions AI can't source are guesses — keep them out of formal material.
The pain of diagramming isn't "can't draw" — it's that edits are exhausting: change one box and you re-layout the whole thing. Generating descriptive, re-flowable diagrams turns "drawing" into "describing structure"; editing text is 10x faster than editing the picture.
The red line for email automation is "draft yes, auto-send careful". "Triage + draft + you hit send" is safest; letting AI blast-send means you own the mistake, not it.
The biggest waste in meetings isn't the meeting — it's nobody remembering "who, by when, what" afterward. The point of this flow is to compress scattered discussion into a task table with owners and due dates, not to write another minutes doc nobody reads.
The most common analytics failure isn't "can't compute" — it's "computed fast but wrong": units not unified, nulls treated as 0,口径 drifting from last time. The point: define口径 first, let AI compute, then spot-check 3 rows by hand.
"Read this PDF" is the easiest request to get wrong. Two reasons: the context window will not hold it, and the model will invent the parts it never saw. The reliable order is extract, then target, then ask, with a page number on every claim.
The classic failure mode is twenty "suggestions" where eighteen are style preferences. A useful review contains only two kinds of information: what will break, and what will confuse the next reader. So the job is not to make AI say more. It is to forbid everything else.
Xiaohongshu posts fail in a specific way: the content is fine and nobody taps. The first decision point is the cover and title, not the body. So reverse the order. Lock the cover line first, then write the content.
The common mistake is exposing your database, schema or API as a pile of tools. You end up with dozens of tools, the model picks the wrong one, and every call fails. Design tools around the handful of things a person would actually ask for. One action, one tool.
The usual failure is not "AI cannot make slides". It is slides where every page is a paragraph: dense text, no focus, and the presenter ends up reading aloud. This workflow splits *what to say* from *how it looks*, with a manual trim in between.
>-