The weekly AI question

One question a week, four options, one best answer, and the page that explains it. New every Monday; written in advance by the chapter, so the date on each is when it was published.

This week

The weekly AI question ·

Monday, again

Every Monday you explain the project to the assistant from scratch - who the sponsor is, what the report looks like, which language - before you can ask for anything. What fixes this?

Pick one, then show the answer
Show the answer

Best answer

BOne project workspace per real project: the charter, the WBS and last report uploaded once, the standing instructions written once.

Four words for four things: a prompt is the ask, custom instructions are how you always want it to behave, a project is a folder that remembers, a skill is a recipe it can run. The one-project-per-project rule puts the project's documents and instructions where they belong, once.

Why, option by option

  1. AIt works, and it is what most people do. But it is the manual version of a feature all four assistants have, and the pasted document counts against the room the conversation has for this week's material.
  2. BThis is what a project workspace is for. Set it up once and every Monday's message is just "here are this week's notes".
  3. CCustom instructions apply to every chat, including the ones about other projects and the recipe you asked for on Sunday. They are for how you always want it to behave, not for one project's facts.
  4. DMemory holds preferences and small facts about you, not a charter. It is also the one place you should be most careful about what goes in.

Read why: Four ways to ask

Earlier questions

The weekly AI question ·

The generic risk register

You ask an assistant to "write a risk register for a software project" and get twelve risks that would fit any project ever run. What is the first thing to change?

Pick one, then show the answer
Show the answer

Best answer

CWhat you give it - upload the charter and the WBS, then ask for risks specific to this project.

The single idea that separates useful answers from fluff: the assistant has read the internet and has not read your project. Context is everything it can see in this conversation - text, files, instructions - and a PM already owns the perfect context documents. The charter and the WBS turn "risks for a software project" into "risks for this one".

Why, option by option

  1. ALonger wording on top of no information still produces a generic answer. The assistant is not being vague; it has nothing specific to be specific about.
  2. BNone of the four knows your project. Each knows exactly what you put in front of it in this conversation, and that is the same nothing.
  3. CThis is context. With the charter and the WBS in front of it, "ten risks specific to this project, each with cause, effect, and an owner role" produces a register you can argue with in the meeting.
  4. DA generic list costs more to make specific than a specific list costs to check. The forty minutes of editing are the reason people give up on the tool.

Read why: Context

The weekly AI question ·

The register in the tab

A colleague has the stakeholder register open - names, phone numbers, who is difficult - and is about to paste it into an AI assistant in a browser tab, signed out, to "make it into a nice table". What do you say first?

Pick one, then show the answer
Show the answer

Best answer

BReplace the names with roles and strip the phone numbers first; then the table works exactly as well.

The test is not "is the AI safe" but "would I email this to a stranger". A stakeholder register with names and numbers is personal data under the Law on Personal Data Protection, and pasting it into a consumer service abroad is processing and a transfer. Most of the jobs a PM hands to an assistant work exactly as well with roles as with names - so the habit to build is the thirty-second anonymiser, before the question of which tool even comes up.

Why, option by option

  1. ASigned out is the state with the fewest data controls, not the most. It is still a third-party service in another country receiving personal data.
  2. BRoles do the job. "The sponsor", "vendor A" and "the site engineer" sort, filter and format like names do, and nothing personal has left the building.
  3. CSigning in gives you controls, and you should - but the register still went out with the names on it. Deleting a chat afterwards is not the same as not sending it.
  4. DThe employer's tool is the right place for work material, and this is the better half of an answer. But "nice table" needs no names at all - anonymising costs thirty seconds and removes the question entirely.

Read why: What never goes in

Last checked by Petar Trajchevski. Product names and settings paths were walked through on that date; the concepts do not move. Report a change