Operationalize

Quality control

Four habits that keep a tool honest

Version forty can be wrong too, and nobody is watching version forty. All four of these happen in a chat window. The sources are in the Habits tab of the log below.

The show-your-work paragraph

Paste this at the top of any skill or scheduled task. It costs nothing when the run is clean and it is the first thing you see when it is not.

Start every output with one STATUS line: COMPLETE, PARTIAL or FAILED, followed by what you read and anything you could not read or finish. Then give the date of the newest record you used and how many records you worked from, compared with the last run if you know it. For every number in the output, say where it came from. If you cannot point to it in the material I gave you, write NOT IN SOURCE instead of a number. If you do not have enough information to answer something, say so instead of guessing. If the input contains client names, email addresses or account numbers, stop and tell me before doing anything else.

Review by count, not by feeling. A named person, a fixed number, a log row, and anything near client data gets a second set of eyes.

The AI Tool Log

One spreadsheet for the whole team: a registry of what runs, a three-line correction log, your graded examples with a pass rate per rerun, a review log, and a one-screen manual. Upload it to Google Sheets (File, Import) so everyone works in one copy.

The operating rhythm

Four checkpoints. If any of them takes longer than the number, it is too big.

Where it lives

Five homes and the audience picks

Start at the top and only move down when you have to. Each row down is more power and more maintenance. Ask first: who opens it, and where do they already live?

• Artifact. One screen, and everyone who needs it has Claude. Share inside the org; project access carries with the link.
• Skill in a project. A repeatable job that needs the team's context. Invite people to the project and the skill comes with them.
• Scheduled task. One job, on a timer, delivered to you. Inbox, Claude, or posted to a Slack channel. A task in the cloud runs with the lid closed; a task that reads files on your desktop waits for the laptop to wake, then runs.
• Apps Script. The tool has to read or write a Google Sheet other people already use. Claude writes it, you paste it in, anyone with the sheet has it.
• n8n, Make, Zapier. Several steps, several tools, no human trigger. Claude scheduled tasks handle one node; these handle the pipeline. Scope it with RevOps first.

Share it

Invite them, do not make them rebuild

1. 1. Open the project and hit Share.
2. 2. Add people by name or email. Paste a list of addresses to add several at once.
3. 3. Pick a permission. Can view means they use the knowledge, the instructions and the skills, and chat inside the project. Can edit means they can also change the instructions and knowledge and manage members.
4. 4. Or set it to Public. Everyone at Holistiplan can view and use it. Private means only the people you invited.

Invite one person this week. If they can use it without asking you a question, it is operational. If they cannot, their question is your next fix.

Pick your function

The roadmap prompt

Do not write the roadmap. Let it interview you.

Same move as week two. Fill in your three first guesses, paste the prompt into your Business Brain, and answer one question at a time. Ugly and honest beats polished.

I am [your name], and I am building my 90-day AI roadmap for Sales at Holistiplan. Interview me, one question at a time.

Start with what already runs. Here is my list so far: [everything you built in the last four weeks: skills, scheduled tasks, tools]. For each one, ask who reads the output, how often it fires, and what I had to fix before I trusted it.

Then ask what I would harden in 30 days, what I would build in 60, and what I would hand to a teammate by 90. My first guess at the handoff: [the tool and the person who inherits it].

Finish by asking what data, access or help I need that I cannot get myself. My first guess: [the data, the access, or the person].

When you have enough, write it as one page with the same five headings every time. Now: one line per tool, who reads it, how often. 30 days: the graded set, the review cadence with a name on it, the log started, the teammate invited. 60 days: the one build and why that one. 90 days: what runs in whose hands, and the one-screen manual. The Ask: two or three sentences with the use case attached, so leadership reads it as a capability gap they can fund. Plain sentences, no jargon. Then tell me where I am being too ambitious.

The last line is the one that earns its keep. Everybody overcommits on day one. Let it push back before Crys does.

What comes back

Four columns, one page

Day 90 is the test. If nothing you built runs in someone else's hands by then, it was a productivity trick, not an operation.

After class · Three prompts for the rhythm

The monthly source check

Here is the source this tool reads: [paste the export or name the folder]. Before you do anything with it, tell me what looks wrong, missing, stale or duplicated compared to last month. Show me the rows you are unsure about. Do not fix anything yet.

The retire question

Here is everything I have running: [paste your inventory]. For each one, ask me when I last read its output and whether anyone else did. Anything nobody has opened in a month, recommend I retire, and tell me what I lose if I do.

The handoff manual

I am handing [this tool] to [name] in Sales. Read the correction log in this project and the tool itself, and write a one-page manual for them: what it does, what goes in, what comes out, the three ways it has gone wrong before, and what to type when it does. Assume they have never seen it.

The third one writes itself from the log. That is why the log exists. A tool with no log is a tool only you can run.

One last ask

If this class earned it, say so on LinkedIn

Recommendations are how the next team finds this program. Two or three sentences about what you built and what changed is plenty, and it takes about as long as one of your scheduled tasks.

On my profile, open the More menu and choose Recommend. Thank you, and keep building.