CCAZ · SESSION 2 OF 2 · BEYOND PROMPTING · SEP 11, 2026

Beyond prompting: turn one use case into a tool your team runs.

Session 1 taught the prompt. This session is what comes after it: identify a use case, give the tool the right sources, build it as a saved agent, refine it until it is right, and know what it takes to connect it to the systems you already run. We do all of it live in Copilot, ChatGPT, and Claude on a construction crew planner. Most of the room has Copilot at work; the other two are where the rest of the room already is. This page is the board for the session: the steps, the copy blocks, and the checks.

Goal 01

Identify a use case worth building, and the sources it needs.

A repeating task, the documents it starts from, and where those documents live. If the tool cannot reach the source, the use case is not ready. That is the first filter, and it takes ten minutes.

Goal 02

Build a saved tool, refine it in rounds, and know how to connect it.

An agent with your rules written in and your files attached, built in three rounds in front of the room. Then the two permissions on every connector, and the one question to ask IT.

01 / 08 · Which model is answering · 5 minutes

Same question, three sizes of model. Stop when the answer is right.

Every tool has more than one model behind it, and the picker is usually one click from the chat box. Small models are fast and cheap and fine for most of the day. They also miss things a bigger model catches. Before we build anything, one question, asked up the ladder until it is answered correctly.

run · copy the question below · claude haiku first, new chat each rung

The question · paste it as written

The right answer is drive. The car has to be at the car wash.

I need to wash my car. The car wash is 40 meters away. Should I walk or drive to get to the car wash?

Claude

1Haiku. The small, fast one. Paste the question. Expect "walk."

2Sonnet. The default. Same question, new chat. It usually catches it. If not, one more.

3Opus. The big one. Stop when it says drive and explains why.

ChatGPT

1The fast or mini option in the model picker, top left. Paste. Expect "walk."

2The default model. New chat, same question.

3Thinking. Slower. Stop when it says drive.

Microsoft 365 Copilot

1Quick response. The default under the chat box. Paste. See what it does.

2Think deeper. Same question. This is the only switch most Copilot users have, and it matters.

3Say what changed. Same tool, same account, same question. Different model, different answer.

TrackingAI.org chart of AI models plotted on a bell curve of average IQ from 50 to 150. Small and free models cluster between 75 and 100. The paid thinking models from OpenAI, Anthropic, Google, and others sit between 120 and 137. — The same gap, measured. This is an offline reasoning quiz the models have not seen in training, averaged over the last seven runs. The free and small models sit in the middle of the curve; the Copilot dot near 85 is Bing Copilot, the free consumer chat, not the Microsoft 365 Copilot on your work account. The paid thinking models sit on the right, and that is where Think deeper, Thinking, and Opus live. When a tool gives you a picker, this is what you are picking between. Two rules for the day: use the default for the routine, and move up one rung any time the answer has to be right. Chart: TrackingAI.org, a MaximumTruth.org project. Captured September 2026. Scores move every month; the shape does not.

02 / 08 · One wireframe, every tool · 5 minutes

One wireframe. Five different brains.

Copilot, ChatGPT, Claude, Gemini, Perplexity: the windows are laid out the same way. The seven numbered areas below are where everything in this session lives. Hover or tap a number to see what that part does and what each tool calls it. The rest of the session is areas 1, 2, 4, and 6.

talk · pin 1, 2, 4, 6 · no files

Model: default ▾

⚙

Ask anything about the look-ahead, the crew, or the schedule. ⏎

Platform features

Connectors

Your mail, files, and systems, read by the tool.

Research

Several minutes, many sources, every one cited.

Files and sheets

Attach a spreadsheet. Work in it.

Voice

Talk to it from the truck.

Hover or tap

Click any numbered area on the wireframe.

Each area explains what that part of the window does and what Copilot, ChatGPT, Claude, Gemini, and Perplexity each call it.

The windows look almost identical. The differences are behind each area, and in which account you are signed into.

03 / 08 · Identify the use case and its sources · 10 minutes

Name the repeating task. Then name where its documents live.

Call out the documents your day starts with: the look-ahead, the schedule, the GC's email, the pay app, the daily report, a text from the field. Every answer goes on the board. Then we sort them by source: email, calendar, storage, spreadsheet, project management, accounting and CRM, texts and phone. A use case is worth building when the task repeats and the tool can reach the source. The sort shows both.

run · type answers into the board · handout

Type an answer, press Enter. Then pick the source it lives in.

0 on the board

Most rooms land heaviest in email and spreadsheets, with the project management system close behind. Notice how many answers are in a system nobody has pointed the tool at yet.

The common sources, and what reaches them today

Outlook or Gmail. The GC's schedule change, the RFI response, the short pay app notice. Most of the day starts here.

Copilot Work IQ reads it

Calendar

Crew starts, inspections, the OAC meeting, the weekly schedule meeting. Dates the tool can only know if it can see the calendar.

Copilot Work IQ reads it

Storage

SharePoint, OneDrive, Google Drive, Egnyte. The look-ahead, the subcontract, the spec PDF, the roster. The biggest pile in the room.

Work IQ reads it, or you attach the file

Spreadsheets

The look-ahead in Excel. The schedule of values. The quantity tracker. Copilot sits inside Excel; ChatGPT reads the file you attach.

Attach it, or Copilot in Excel

Project management

Procore, Autodesk Build, Bluebeam. The CO log, RFIs, submittals, daily reports. Where the job actually lives, and where the tool mostly cannot see yet.

Connector, if IT turns one on · Section 06

Accounting and CRM

Sage, Viewpoint, QuickBooks, HubSpot. Pay apps, job cost, who you bid to last year. Same story: a connector, or an export you attach.

Connector or export · Section 06

Texts and phone

The foreman's text at 6:12a. The verbal at the trailer. No tool reaches these. You paste it, or it never happened.

Paste it · Session 1

The sort answers the use-case question. Reachable source plus repeating task equals build it. Unreachable source equals paste it for now, and put it on the connector list for Section 06.

04 / 08 · Give the tool the right context · 12 minutes

Three ways to give it context: Work IQ, an attached file, a saved project.

The same question in Copilot, ChatGPT, and Claude, with the context supplied three different ways. The question: "Which days next week are short on crew, and by how many? Show the math." The steps below are the checklist so nothing gets skipped.

run · lookahead + roster.xlsx / terms.txt · start research in a second tab · all docs

Copilot · Work IQ

ask → reads your mail, files, chats → answer with links

First move: Work IQ on, ask "what did the GC send me this week about schedule changes." It should point at the email. Then turn Work IQ off and ask the same thing. It guesses about GCs in general. Same question, different context.

ChatGPT · a Project

new project → three files in → ask without pasting

Name it after the job. Drop in the look-ahead, the roster, and the terms. Ask the crew question. Then ask a second question with no setup. The context is in the project now, not in your typing.

Claude · a Project

new project → same three files → same two questions

Identical moves, different vendor. Project knowledge holds the files, project instructions hold the standing rules. Show it for sixty seconds so the room sees the idea is not a ChatGPT feature. It is how every one of these tools works.

Checklist · do not skip a step

Four moves in Copilot, four in ChatGPT, one minute in Claude.

Copilot 1

Work IQ on, then off

Same question with Work IQ on, then off. Point at where each answer came from.

Copilot 2

Attach the look-ahead

Upload pour_lookahead_and_roster.xlsx. Ask the crew question. Make it show the math.

Copilot 3

Open it in Excel

Same file, Copilot in the ribbon. "Add a column: crew needed that day across all jobs." Click the formula.

Copilot 4

Check one thing you know

In the sample, 11/05 needs 11 finishers and four are out or on light duty. Short 1. Does its number match? If not, fix the context, not the answer.

ChatGPT 1

Make the project

New project, job name. Three files in.

ChatGPT 2

Paste the standing rules

The house rules from the terms file go into the project instructions. Once.

ChatGPT 3

Ask without pasting

The crew question. Then "which equipment is not booked yet." No setup on the second one.

ChatGPT 4

Say the difference out loud

Chat forgets. Project remembers. Work IQ already knows. Point back at the board: which sources did we just reach, and how.

Claude

Same project, sixty seconds

New project, same files in project knowledge, same house rules in project instructions, the crew question. Same answer shape. The vendor is not the lesson.

If the wifi dies · what the screen should look like (Microsoft 365 Copilot, work account)

Work IQ on. The greeting is about your calendar because it can see it. Researcher and Analyst pinned, Agents below, New agent at the bottom of the list.

Work IQ off. Same window, now a general chatbot: "What can I help with?" Same question, different pile. This is the switch.

The plus button: upload files, attach cloud files, change data sources. This is how a spreadsheet or a PDF from storage gets in without a connector.

Notebooks on the work account. Create new. One per job. The Open in OneNote button is where the notes end up.

If the wifi dies · Notebooks, step by step

Name it the job name. We used Building 3 tilt-up.

Add references: upload, links, or OneDrive. The look-ahead, the roster, the terms.

The notebook, with References on the right. Every chat inside it sees those files. The storage pile, saved once.

Researcher. "What do you want to research today?" with a Sources button. Paste the Section 07 prompt here and let it run while we build.

05 / 08 · Build the tool, then refine it · 28 minutes

The Crew Planner. Built as an agent in three rounds.

An agent is three things: the instructions, the files it reads, and where it stops. In Copilot it is an agent you create from the chat window. In ChatGPT it is a custom GPT. In Claude it is a skill, or a project with the instructions pinned. We build it in Copilot in three rounds, then paste the finished version into the other two. Round one is deliberately thin so the room can catch what it gets wrong. Every catch becomes one line in the instructions. The sample is a concrete crew's look-ahead; the method is the same for any trade.

run · v1 / v2 / v3 · then disruption email · same two files as knowledge

If the wifi dies · the agent builder, three screens

New agent. Either describe it in the box or hit Skip and fill the form yourself. We fill the form; the room needs to see the parts.

Configure: name, then Instructions. This box is where Round 1, 2, and 3 get pasted. Nothing else changes between rounds.

Knowledge: Add knowledge is the two files. Below it, Work content: Cloud files, Outlook, Teams, Copilot connectors, Web search, each with its own switch. All off for the planner.

The disruption · after Round 3

Friday 3:48p. The GC moves a pour two days early.

Paste the email below into the planner and ask: "What breaks if we say yes?" A good answer names the crew shortage on 11/10, the equipment that is not booked and is already inside its booking window, and the material order deadline. It does not say yes or no. That stays with the PM.

Get

The email, as is

Paste it. Do not summarize it. The planner already has the look-ahead and the roster.

Move

Rerun Tuesday 11/10

Mesa slab east, 9 finishers and a booked pump, plus panels 13-18, 6 finishers and a pump. 15 needed, 13 available. Two pumps, one booked.

Decide

The PM answers the GC

Yes with conditions, or no with the reason. The tool gave you the arithmetic and the deadlines. It did not book a pump or move a pour.

Save

One more line

Whatever the room caught this round goes into the instructions. Round 4 happens Monday, at your company, on your look-ahead.

The disruption email · paste into the planner

Sample only. The pump window and the finisher count are the lesson.

From: GC scheduler, Building 3
To: Concrete PM
Sent: Fri 11/06, 3:48p
Subject: Panel pour 13-18 moving up

Erector wants panels 13-18 cast earlier so they can start the north wall the week of 11/16.
Can you cast panels 13-18 on Tue 11/10 instead of Thu 11/12? Forms would need to be done Monday.
Everything else on the look-ahead stays. Let me know Monday morning.

Forms for 13-18 are already on the look-ahead for Monday 11/09, so the forming is not the problem. Finishers and the pump are.

Files it reads

pour_lookahead_and_roster.xlsx (two sheets), ready_mix_and_pump_terms.txt. All three in the downloads on the resources page. Attach them as the agent's knowledge, or put them in the project.

Where it lives

Copilot: create an agent from the chat window, paste the instructions, attach the files. ChatGPT: a custom GPT, same instructions, same files. Claude: a skill, or a project with the instructions pinned and the files in project knowledge. Same three parts every time. The Round 3 text on this page pastes into all three without a word changed.

Where it stops

It plans. It flags. It asks. It does not move a date, book equipment, or send anything. That line is in the instructions on purpose and it is the line you never delete.

06 / 08 · Connect it: two switches on every connector · 8 minutes

Connectors: what the tool may read, and what it may change.

A connector is how the tool reaches a system you run on: Outlook, SharePoint, Procore, your accounting system. Work IQ was one; you used it an hour ago. Every connector has two permissions, and the whole conversation with IT is about which one is on.

talk · screenshots below if needed · no files

Switch one · look

It can read what you can read.

Outlook, Teams, SharePoint, the look-ahead on OneDrive. Same permissions you already have. If you cannot open the file, neither can it. Turn this on first. Watch it for two weeks.

Switch two · change

It can send, file, post, or update.

Leave this off until you have watched it be right on your own files. One large Phoenix contractor runs its Microsoft connector read-only on purpose: it can read Outlook, it cannot draft into it. That is the setting most companies should start with.

The sentence for IT

"Which connectors are on, and are they read-only?"

That is the whole ask. Second question if you have the paid Copilot: "Do I have a Microsoft 365 Copilot license, or just Copilot chat?" Notebooks, Researcher, and agents need the paid one.

If the wifi dies · the two switches, in the product

Every row under Work content is a connector with a switch. That is switch one, look. Off by default. Turn on only what this agent needs to read.

The Cowork tab: tasks that run on their own, Automations, a monthly limit set by the organization. That is switch two, change. It exists. It is not for today.

Watch for this

It can only see what the permissions say, not what is on your screen.

An assistant at that same contractor had full rights to an executive's mailbox and the tool still could not see it, because formal delegate access was not set. If it says it cannot find something you can see, the answer is usually permissions, not the prompt.

Watch for this

The add-in is not in the ribbon.

Copilot in Word and Outlook shows up only if IT turned it on. If you do not see it, that is a ticket, not a training problem.

Still true

The owner's documents stay in the company account.

Work IQ respects the confidentiality clause because it respects the file permissions. A personal ChatGPT account does not know the clause exists.

Not today

Procore, Sage, Bluebeam connectors.

They exist or are coming, and the same two switches apply. Bring the question to a Session 3 if the association wants one: one task, connectors on, look only, in the room.

07 / 08 · Research, and the phrases that keep it honest · 10 minutes

Research: when the answer is outside your files.

Research is for questions about the world, not about your files. Copilot calls it Researcher, ChatGPT calls it deep research, Claude calls it Research. All three need the paid plan. It reads for several minutes and comes back with sources. The output is not on this page on purpose. It comes from the sources on the day.

run · switch to the research tab from 04 · click one source · then the table

Research prompt · start in Section 04

The Task asks for the pages to open, not a conclusion.

Context: I am the PM for a concrete subcontractor in Arizona. A GC superintendent
verbally directed extra work and nothing is in writing yet. Our subcontract has a
written-direction clause and a notice period. I need to know what the law and the
standard subcontract forms say about preserving the claim.

Role: Act as a construction attorney's paralegal who has pulled this research many
times and knows the answer is only useful with the section cited.

Task: Research and report 1) what written notice Arizona law expects a subcontractor
to give to preserve a change order or delay claim, and any statutory deadlines
2) what the common subcontract forms (AIA, ConsensusDocs) say about notice
3) for every claim, the section or source, with a link I can open 4) what I should
confirm against my own subcontract before relying on any of it 5) stop. Do not give
me legal advice. Give me the pages to open.

Three moves when it is back: read the summary out loud, click one source and check the quoted section is there, ask the room whether it matches what they know. If someone disagrees, that is the best five minutes of the day.

Prompts that quietly do the most work.

Short add-on phrases you tack onto any prompt, in any of the tools. Each one changes what the model does behind the scenes. The research prompt above already uses three of them.

What you add to the prompt	What AI does differently
"Show me your citations"	Forces AI to attach a source URL to every claim. If it can't, the gap becomes visible.
"Double-check your work"	AI re-evaluates its own answer. Often catches its own errors before you do.
"Think carefully"	Triggers longer, more deliberate reasoning. Better on multi-step questions.
"Show me only facts, no assumptions"	Reduces speculation. AI sticks to what it can document.
"If you can't find an authoritative source, say so"	Prevents fabrication. AI will admit when it doesn't know.
"Quote the exact passage"	No paraphrase. AI returns the actual words from the source.
"What's the date on this source?"	Currency check. Catches outdated info you'd miss otherwise.
"What are you uncertain about?"	AI flags its own confidence gaps. Tells you what to verify.
"What would I need to verify before acting?"	AI generates a verification checklist. Built-in safety net.
"Compare across multiple sources"	Triangulates instead of trusting one. Surfaces disagreements between sources.

How to use this table

Stack two or three. "Show me your citations, double-check your work, and tell me what I should verify." Each phrase adds a layer of discipline. The combination does more than any one of them, and the lines belong in the Crew Planner's instructions as much as in a research prompt.

08 / 08 · Scale it: your use case, then the month · 12 minutes

Pick two use cases from the board. Start one.

Back to the board in Section 03. Pick two answers. For each: which source is it in, can the tool reach that source today, what does it turn into, where does a person look before it goes out. Then we start a project on screen for one of them, with the owner's name stripped. Ten minutes, two people, one project.

run · back to the board · start one project live · all docs

Whose use case? The look-ahead, the daily report, the pay app, the submittal log. Call it.

This week

Make one project or notebook for the job eating your week. Put the three files you open most in it.
Create the Crew Planner in whichever tool you have: a Copilot agent, a custom GPT, or a Claude skill. Paste the Round 3 instructions. Attach your own look-ahead and roster.
Ask IT the one sentence: which connectors are on, and are they read-only.

Every Monday for a month

Run the planner on the real look-ahead before the schedule meeting.
Write down what it got wrong. Add one line to the instructions. That is Round 4, 5, 6, 7.
Count the minutes the schedule meeting took before and after.

What you skip

No Work IQ off for questions about your own job. Check the switch.
No owner documents in a personal account.
No change permission on any connector. Look only, until you have watched it.

At the end of the month

Show the planner to one other person in your company.
Decide: keep it, change it, or drop it. Any of the three is a good answer.
Bring the number and the instructions file if there is a Session 3.

The same four ideas in three tools.

The idea	Microsoft 365 Copilot	ChatGPT	Claude
A folder with memory	Notebook	Project	Project
A saved way of working (a skill)	Agent, created from the chat window	Custom GPT	Skill, or project instructions
The model picker	Quick response or Think deeper	Model menu, top left	Model menu, under the chat box
A connector	Work IQ; Microsoft 365 built in	Connectors in settings	Connectors in settings
Research with sources	Researcher	Deep research	Research

Product names as of the session date. The free Copilot chat does not include Notebooks, Researcher, or agents.

Common questions

You're probably wondering one of these.

Click a question.

Click a question to see the answer.