Where we are - week 18, and it is not going well
Here is the honest state of Northwind. M3, raw ingest signed off, cleared in week 13 - the gate was originally 10 August, moved to 24 August, and landed later still. M4, the modelled warehouse, is running with one data engineer instead of two: R-03 is half closed, one hire has accepted and one requisition is still open. R-11 is live, with data quality at 96.4% against a 99% gate, rising but not fast enough to be boring. R-07 is quiet, which with Veridian is not the same as fine. Nine source systems, 812k rows a night, and a sponsor who has stopped asking how it is going and started asking whether the Christmas date holds. That is a different question and it deserves a different kind of answer. Tonight you run the week Monday to Friday using everything at once: the workspace from b1, the charter from b2, the milestone architecture from b3, the dependencies from b4, the register from b5, the six checks from b6, the decision log from b7, the escalation craft from b8 and the connected pack from b9. Five real artifacts and one plan for yourself. Nothing new gets introduced, which is the point of a capstone.
Monday to Friday, and where you must be in the room 7 min live
A delivery week has a shape whether or not you have named it. Most weeks go wrong in the same place: the report gets written on Thursday morning from whatever happened to be visible, so the gaps get filled by inference instead of by asking people. Moving the draft to Wednesday changes that, because it leaves you a day to go and get the real answers. That single shift is most of the value in this map.
LiveThe five days, one paragraph each4 min▶
- Monday - refresh the pack and the register. The connected pack pulls the week's movement and diffs it. You read the diff and confirm it is true, because a diff can be technically correct and materially wrong: a ticket closed on Friday afternoon that was not really done is a lie the pack will happily repeat all week.
- Tuesday - dependency and risk sweep. What moved, what is now blocking what, and which risks changed state. For Northwind week 18 that is the M4 critical path with one engineer, R-03's open req, and whether R-11's 96.4% moved. Then you confirm the important ones with the people who own them. Ten minutes on the phone beats an hour of inference.
- Wednesday - draft the report and go and fill the gaps. The report drafts against the six checks with its GAPS block on top. Then you leave your desk. Every gap gets filled by a person: Marcus on the modelling estimate, Priya on Veridian, Sofia on the schema decision, Dan on BI readiness. This is the day that separates a report a steerco can act on from one it can only read.
- Thursday - steerco, and decisions logged the same day. You run it. The pack drafts the decision capture from your notes within the hour, while you still remember what was actually meant rather than what was said. Same-day logging is not tidiness - a decision recorded on Monday is a decision reconstructed.
- Friday - actions out, log updated, next week parked. Actions go out with one owner and one date each. The decision log is updated and visible. Anything you could not answer this week gets written down as next week's opening question rather than carried in your head over the weekend.
The Wednesday that saved the quarter. A delivery lead moved her draft from Thursday 7am to Wednesday 10am and did nothing else differently. That week the GAPS block asked who owned the migration cutover, and because it was Wednesday she had time to walk over and ask. The answer was that nobody did - the person everyone assumed owned it had moved teams in July. On the old schedule she would have written "platform team" at 7:40am on Thursday and it would have gone to the sponsor as a fact. One day of slack turned an invented owner into a real finding.
Self-studyAdapting the map to your cadence2 min read▶
The days are not sacred. The sequence is. Whatever your week looks like, keep these four properties and it works:
- The refresh comes before anything is drafted. Drafting from a stale pack produces confident answers about last month.
- There is a full working day between the draft and the meeting. That day is not slack, it is the gap-filling day. Cut it and you have automated inference.
- Decisions are logged the same day they are taken. Not "by Friday". The same day.
- Actions leave the building with one owner and one date. If you cannot name the owner on Friday, that is the finding, and it goes in the report rather than being smoothed over.
Fortnightly steerco? Run the same sequence on a two-week rhythm and keep the gap-filling day. Daily-ship team with no steerco at all? The report becomes a Friday summary and the decision log matters more, not less, because nobody is in a room together to remember.
Run the week: five prompts, all of them short 6 min live
Notice how small these are. That is the point, and it is rail one arriving at its destination: the workspace already knows Northwind, the skill already knows your format, the connector already has this week's state. There is nothing left to explain. A long prompt at this stage usually means something is missing from the workspace and you are trying to paper over it with words.
Monday, 09:00. Refresh the pack and diff it. Read the diff yourself before anything else touches it, and challenge one line of it out loud.
Tuesday, 09:00. Sweep dependencies and risks against the M4 critical path. Then call the two people whose answers would change the report.
Wednesday, 10:00 and again at 16:00. Draft the report with the GAPS block on top, spend the afternoon away from your desk closing the gaps with humans, then re-run with them filled. Read every number. Send the pack to attendees the evening before, never in the meeting.
Thursday, immediately after steerco. Capture decisions from your notes while the meaning is still fresh, and log them today.
Friday, 14:00. Actions out with owners and dates, decision log published, next week's open questions parked in the workspace.
Two contract engineers, and the number that made the decision easy. The Northwind ask in week 18 came down to one option with a price: two contract engineers for four weeks, roughly £38k, to hold the M4 date while R-03's open req runs its course. The draft laid out three options with costs and an expiry date. Elena took ninety seconds. What made it fast was not the writing - it was that the number had a source, the date had come from Marcus, and the trade-off had been named rather than implied. A sponsor cannot decide quickly on a paragraph of texture.
Four things you never delegate, and why 6 min live
Ten sessions of moving work off your desk earns you the right to be precise about what stays on it. This is not a safety disclaimer and it is not modesty about the technology. Each of these four is a place where being accountable is the work, and the model has nothing at stake - no reputation to lose, no relationship to repair, nothing that happens to it if it is wrong.
LiveWhy each one stays with you4 min▶
- Commitments. A date is a promise made on behalf of people who will be held to it. Its value comes entirely from the fact that someone with something to lose checked with them and then said it. A generated date has the shape of a commitment and none of the substance, which is worse than no date, because it will be planned against. This is rail two, and it is the rail people break first when the pressure is on.
- Bad news. The message that costs you something to send is the one that buys trust. When you tell Elena the Christmas date is at risk and you are the one who says it, the relationship survives the news. Delegate the drafting of the words if it helps you find them - but the sending, the tone in the room, and the follow-up question you did not anticipate are yours. Session b8 was about doing this well; tonight is about not outsourcing it.
- People. Performance, morale, someone who is quietly about to resign. These are judgements about human beings made with incomplete and often deliberately private information, and the model is missing every signal that matters: how Marcus sounded on Tuesday, that Dan has stopped arguing in meetings, that Sofia asked an oddly specific question about handover. It is also the fastest way to do real harm with a tool. Do not put a person's name into a workspace attached to a judgement about them.
- The trade-off call. Scope, cost or date - one of them moves. You can and should have the options and their costs drafted; that is genuinely useful and this course has taught you how. Choosing is a different act. It requires knowing what your organisation will forgive, what your sponsor said in a corridor in March, and what you are willing to be wrong about in public. Somebody has to choose, and it has to be somebody who lives with it.
The polished escalation nobody believed. A PM sent a well-structured escalation about a slipping vendor - clear, evidenced, correctly formatted, entirely drafted and barely read. In the meeting the sponsor asked one follow-up: "how bad does this get if they miss again in October?" He did not know, because he had never sat with the question long enough to have a view. The escalation was correct and it landed as noise, because the person sending it had not paid the cost of thinking it through. The artifact was fine. The accountability had not happened.
Self-studyThe grey zone, and a test for it2 min read▶
Plenty of work sits between "obviously draftable" and "obviously mine". A reasonable test: if this turns out to be wrong, does someone need a person to be answerable for it? Drafting the options and costs for a trade-off passes - draft it, you still choose. Writing the words of a hard message passes, as long as you rewrite it in your voice and send it yourself. Proposing a date from the plan structure does not: draft the structure, get the date from the team. Summarising a one-to-one with a struggling engineer does not go anywhere near a workspace at all.
When you genuinely cannot tell, assume it is yours. The cost of doing something manually that could have been drafted is twenty minutes. The cost of the reverse can be a programme.
Run Northwind's recovery week, end to end ★ 16 min · everyone builds
Five artifacts, in order, on the hardest week this programme has had. Work in your Northwind workspace with the pack loaded. If you are running this solo after the session, give yourself the full sixteen minutes and do not skip the gap-filling step just because there is nobody to ring - write down who you would have called and what you would have asked.
Artifact 1 - the refreshed pack. Run Monday's diff prompt. Output: three lists plus one throughput number, everything cited. Then challenge one line of it. On Northwind week 18 the line worth challenging is any ticket that closed without a linked test result, because M3's late sign-off came from exactly that pattern.
Artifact 2 - the updated risk register. R-03: one engineer accepted, one req open, and it is still a hard gate before M4 completes. R-07: Veridian quiet, which needs a check with Priya rather than a green. R-11: 96.4% against 99%, rising - is the trend real or is it three good nights? Update states, triggers and owners, and note what would have to change for each to close.
Artifact 3 - the report with its gaps block. Run Wednesday's prompt. Read the GAPS block before the report, as always. Expect gaps around the M4 modelling estimate with one engineer, the second hire's start date, and whether Dan's BI team can absorb a compressed M5 window. Assign each gap to a name from the owner list, then fill them - from people, not from the model.
Artifact 4 - the decision log entry. Elena decides on the contract engineers: two for four weeks, roughly £38k, to hold the M4 date. Capture it same-day with id, date, who took it, what it costs, and what it now depends on. Then write the "not actually decided" list - on this week that is usually the Christmas date itself, which people will feel was confirmed and was not.
Artifact 5 - Friday's action list. One owner, one date, each. Marcus on the revised M4 estimate with contractors landed. Priya on the Veridian check-in and VN-2291. Sofia on the schema decision that unblocks modelling. Dan on the M5 readiness view. If any owner is "the team", you have not finished.
The last step, and it is not optional. Read all five as a set, as Elena would. Do they tell one consistent story? A common capstone finding: the register says amber, the report says the date holds, and the action list quietly assumes contractors who have not been approved yet. Fix the story, not the individual documents.
Your own thirty-day adoption plan ★ 8 min · build your own
The failure mode after a course like this is doing all of it for one enthusiastic week and none of it by October. So: one thing per week, in an order where each week makes the next one easier, and one measure so you can tell in a month whether it worked rather than whether it felt good.
Week 1 - the workspace. One programme, six documents, house rules written. Nothing clever. If you only ever do week 1, you have still moved the needle, because context is where the value is.
Week 2 - one artifact. Pick the thing you dread most, which for most people is the status report. Draft it there every week for a fortnight. Do not add a second artifact until the first one is genuinely habit.
Week 3 - the rubric. Score what comes out against the six checks from b6. This is the week you stop being impressed by fluent output and start being able to tell good from plausible, which is the actual skill this course teaches.
Week 4 - the connected pack. One connector, read-only, scoped. The Thursday run if your plan allows it. By now you know exactly what you want it to produce, which is why this is week 4 and not week 1. And pick your measure before you start, written down where you will find it in a month.
The PMO that measured the wrong thing for a quarter. A delivery function rolled this out and tracked hours saved. The number looked excellent and nobody's programme ran any better, because the hours saved went straight into more meetings. The following quarter they switched to decision latency: days from raising an ask to getting an answer. That number was ugly at first - eleven days - and it was the first metric anyone had ever produced that made a sponsor change their own behaviour. Measure the thing you actually want to move.
Ten sessions, three rails, one job that is still yours 4 min live
You started with a mandate email and a blank workspace. You now have a charter, a milestone architecture with gate criteria, a dependency map, a self-maintaining risk register, a report that survives a steerco, a decision log, an escalation you can defend, a connected weekly pack, and a full delivery week you have run end to end. Almost none of that was about prompting. Nearly all of it was about context, structure, and knowing what good looks like before you ask for it.
LiveThe three rails, restated for the last time2 min▶
- Context beats prompting. When a draft comes back generic, the fix is a missing document, not a cleverer sentence. Every good week in this course started with a pack that was current.
- The dates come from the team. Draft the structure of a plan - the milestones, the gates, the dependencies, the questions to ask. Get the dates from the people who will be held to them. A generated date is fiction with good typography.
- A named human signs it. Nothing sends itself. Read it, source every number, own the consequence. "The AI wrote it" has never worked at a steerco and it never will.
Techniques will change under you. The tooling layer in b9 will look different within a year. These three will not, because each one is about where accountability lives, and that has not moved at all.
Self-studyWhere to go next2 min read▶
- If your next problem is organisational, take the leader track. Sessions a1 to a6 cover what AI changes about delivery at a function level, the governed setup and the permission policy that b9 kept handing off, a review rubric your PMO can apply to anyone's AI-drafted plan, escalation you can defend, the reporting cadence, and how to run a pilot with a real quality bar. It is written for heads of delivery and PMO leads - and for you, the week you are asked "should we roll this out".
- If you want the artifacts without any AI at all, take the prequel. learn-tech-project-pmo-with-phoebe builds the same Northwind programme by hand: charter, milestones, dependencies, RACI, risk register, tracker. Nothing generated, everything reasoned. It is the best possible companion to this course, and if you are teaching a junior PM, start them there - you cannot review a draft well if you have never built the thing yourself.
- If a rail broke this month, go back one session. Generic drafts means b1. Dates you cannot defend means b3 and b4. A report that gets nodded at means b6. An escalation that landed badly means b8. The track is designed to be re-entered, not just finished.
One last thing. The reason this course spent ten sessions on artifacts rather than on prompting is that the artifacts were never the job either. The report was always a way of making a programme legible enough for someone to decide something. Now that the drafting is cheap, the question is what you do with the hours - and the honest answer, from every delivery lead who has actually made this stick, is: talk to more people. That is the part nothing has automated.
Try it yourself - the first thirty days ◐ 30-45 min to set up
- Run one full week on your own programme, not Northwind. Five artifacts, in order, with the gap-filling day protected in your calendar so it does not get eaten.
- Write your thirty-day plan on one page: week 1 workspace, week 2 one artifact, week 3 the rubric, week 4 the connected pack. Put it somewhere you will see it, and record your measure's baseline today - decision latency if you are brave, six-check score if you want the most useful signal, hours if you need to show someone a number.
- Write your own four never-delegate lines in your words, and share them with your team. Rolling this out without that list is how a team ends up with a generated commitment in front of a sponsor.
- Book fifteen minutes with whoever governs tooling at your organisation before you scale any of this past yourself. Leader session a2 is the agenda for that meeting.
Official sources covered
The capstone draws on everything the track was built from: the delivery canon for the week's shape, PMI's AI standard for the accountability boundary, the by-hand prequel for the artifacts themselves, and vendor documentation for the tooling that assembles them. PMI certification and the full normative text of the standard stay with PMI. This session covers:
Three questions before you go 🎯 ◐ 90 seconds
1 · Why does the report get drafted on Wednesday rather than Thursday morning?
The gap-filling day is the whole point of the shift. Draft at 7am on the day of the steerco and every unknown gets smoothed over by inference, which is exactly the failure the GAPS block exists to expose.
2 · Marcus gives you a revised M4 estimate. The draft plan already contains a date two weeks earlier. Which goes in the report?
Rail two, at the moment it is hardest to hold. The drafted date was never checked with anyone and will be planned against as if it were. A midpoint is worse still: it belongs to nobody and nobody will defend it.
3 · Which of these four is a genuine never-delegate, and what makes it one?
Options and action lists are draftable and should be drafted. Bad news is not: the value of the message comes from a person with something at stake choosing to say it, and from being in the room for the question you did not anticipate.