Free toolkit · The AI Toolkit

Scope creep never arrives as a variation.

It arrives on a site walk, in an email, and in a conversation in the yard that nobody wrote down. This toolkit sets up an AI workflow that reads a project folder and tells you what changed that nobody approved, in Claude, ChatGPT or Microsoft Copilot. It drafts. A person decides. It never contacts a client.

Get the toolkit → See what is inside

8 judgement checkpoints 3 platform tracks 1 demo project with planted findings 10 staff SOPs

What it answers

Four questions a project manager never has time to answer by hand.

The evidence is already in the project: the signed scope, the meeting transcripts, the emails, the decisions log. Nobody has the hours to read all of it against the scope every week. An AI tool can, once you have told it what good looks like on your projects.

SCOPE

What has changed that nobody approved?

Read against the signed scope. A change only counts as approved when it has an approver and a date.

DECISIONS

What did we decide that never got written down?

Decisions made on a site walk or in a reply to an email, and missing from the log.

AUTHORITY

Who has gone quiet who is supposed to be deciding?

The sponsor who stopped coming to meetings, and the site contact who now relays for them.

IMPACT

What does that do to the date, the money and the promise?

The buffer to a hard deadline, the hours nobody is billing, and the deliverable quietly made smaller.

What is inside

Decide, build, then test it on a project with problems planted in it.

Everything you need to set it up this week, and the governance to run it on live projects without regretting it.

The decisions and the standard.

The part only your team can write. Eight judgement calls made by the people who run the projects, and the standard every finding is judged against.

Eight judgement checkpoints, each with a default
A builder page that writes the instructions as you decide
The same decisions as a spreadsheet
What good looks like, with Green, Amber and Red anchors
Ten staff SOPs and a folder structure

The agent and the platform tracks.

Build it in the tool your business already pays for. The setup differs between the three. The method does not.

Setup guides for Claude, ChatGPT and Microsoft Copilot
The instructions, plain version 1 and hardened version 2
The prompt set, and a run and alert spec
What Copilot can and cannot see in a folder
Why we design in Claude and deploy in Copilot, as a PDF

A demo project and the governance.

A complete fictional project with eight problems planted in it, so you can test your build before it goes near real work.

Twelve project files: SOW, plan, transcripts, emails, logs
Eight findings to catch, and one approved variation for contrast
What a good run finds, to mark yourself against
A permissions trap: another client's file in the same library
Failure modes, an AI use policy and an agent register

Three platforms

Build it where your projects already live.

The decisions, the standard and the instructions are the same in all three. Only the setup differs, and each track is written out step by step.

	Microsoft 365 Copilot	Claude	ChatGPT
Best when	Your projects live in SharePoint, Teams and Outlook.	You want the closest reading of long transcripts, and you are happy to attach files.	Your team already works in ChatGPT.
What it reads	Your project folder in place, and only what the person running it can already open. Nothing is uploaded.	What you attach. The standard sits in the Project knowledge. Its Microsoft 365 connector can also search SharePoint, once your Microsoft admin approves it.	What you attach. The standard sits with the Project. Its SharePoint app can reach your files, where your plan and Microsoft admin allow it.
Set-up	An agent grounded on the project folder. About 30 minutes, plus a few minutes for new files to be ready.	A Project, the standard as knowledge, paste the instructions. About 15 minutes.	A Project with the standard and the instructions. About 15 minutes.
Watch for	It shows whatever the folder shows. Check who can open it before you share the agent.	A file you forgot to attach reads as a quiet week.	Same as Claude. Build it as a Project, not a custom GPT: OpenAI retires custom GPTs on 11 December 2026.

Whichever you pick, the agent only sees what you give it. Copilot cannot read a colleague's mailbox or a transcript of a meeting they did not attend. That is why the pack's staff SOPs matter more than the prompt: what reaches the folder decides what the agent can find.

Before anyone trusts it

Test it on a project with eight things wrong.

Kōwhai Energy, an invented solar installer, is fitting panels at Te Awa Estate, an invented winery, with crews off the roof by 1 November for vintage. One variation was raised and approved properly. Everything else that changed was agreed on a site walk, in an email, or in a conversation that only made it into a Friday file note.

A first run with a plain prompt usually finds three or four of the eight. That is the point of the exercise. The pack shows you how to harden the instructions until it finds them all, and what a good run looks like.

The rules that make it trustworthy

The signed scope is the baseline. Only a change with an approver and a date is approved.
Evidence or it did not happen. Every finding cites a file and a date.
Never invent a number. "Not priced" is the right answer when nothing is written down.
Not found is a finding. A missing source lowers confidence. It never means all clear.
Process, not people. Describe the gap, never name a culprit.
A human reads every output before any of it reaches a client or a conversation.

The 30 minute set-up

Time	Step
5 min	Read the README to the end.
10 min	Make the eight decisions, in the builder page or on paper.
10 min	Make the Project Standard yours: your anchors, your hard rules.
5 to 30 min	Set it up in Claude, ChatGPT or Copilot.
10 min	Run it on the demo project before you point it at real work.
Ongoing	Agree the SOPs with your team.

Read this before you trust a green status

Five ways it goes wrong.

The empty folder. It reports a clean project while the worst item sits in someone's inbox.

False approval. It calls something approved because the client said yes in an email.

Invented numbers. A tidy cost estimate appears that nobody ever wrote.

Blame drift. The output starts to read like a case against the project manager.

Green because quiet. A project goes silent, and silence looks like health.

Get the toolkit

The Project Activity Tracker Toolkit, free.

The eight judgement checkpoints with a builder page and a spreadsheet version, the Project Standard, ten staff SOPs, setup guides for Claude, ChatGPT and Copilot, the instructions and prompts, a complete demo project with an answer key, and the governance templates. ZIP, about 500 KB.