AI Prompts & Workflows Archives - Page 2 of 33

Connect your sheet and GPT with a working trigger

If you’re trying to push a weekly report out of Google Sheets using GPT, the first real hurdle is just getting a reliable trigger to fire. I’ve had triggers that would silently fail without any error messages, especially when I used Google Apps Script time-based triggers. It’s oddly common for these to get stuck or never even run, particularly when Sheets hasn’t been opened recently.

So here’s what I switched to: I created a Zapier automation (Zap #17, currently labeled “GPT Weekly Mess v3,” if you must know 😅) that runs every Monday at 9 AM. The trigger is ‘Schedule by Zapier’ — which in my experience is weirdly more dependable for weekly timing than writing a time-based trigger in Apps Script.

The sheet itself contains a list of metrics from different teams: support ticket count, average response time, NPS score, revenue data from Stripe, and churn percentages pulled via Supermetrics. All of it gets dropped into one master sheet. To avoid API limits or random timeout errors, I make sure the Zap only reads the last row — no full-sheet scans, or you’re begging for the Google Sheets API to throttle you.

On the Zap trigger, I added a lookup formatter step that digs into the last seven rows and checks for a column labeled ‘Report Status’. If it’s marked ‘READY’, we proceed. If it’s blank, the Zap ends silently. This saved me from getting half-built draft reports sent to Slack, which I definitely learned the hard way 🙃

Clean and prep your data before asking anything

If you just dump raw Google Sheets data into GPT with no cleanup, the results are either hilariously wrong or totally useless. I had a row that said “3.2%⬆️” for churn and GPT interpreted that as “churn decreased by three percent” — which is the opposite of what it meant. So yeah, emojis don’t help here 👀

Now I run a pre-cleaning function using Google Apps Script that formats everything into consistent language. Here’s roughly what happens:

– Percent signs get stripped
– “⬆️” and “⬇️” are replaced with words like “increased” or “decreased”
– Decimal formats that don’t make sense (like .9 being interpreted weirdly) are bumped to readable percentages
– Empty or null cells get replaced with placeholders like “No data reported”

You can do this with a built-in formula too, but I found it easier to keep all logic in one script tied to a ‘CleanData’ button inside the Sheet that I can click manually. Was supposed to auto-run but… yeah never dependable 😑

After cleaning, I concatenate the rows into a big block of text using `JOIN()` across a helper column. This is what gets fed into GPT. Think of it like creating a big paragraph summarizing your metrics — GPT tends to behave better with that than raw tables.

Use GPT to prep draft text not conclusions

Let’s talk prompt design. If your prompt sounds like, “Summarize the following KPIs and conclude whether performance was good or bad,” you’ll get AI hallucinations. I’ve had GPT say “Churn improved across the board” when churn actually spiked — it just didn’t understand where the line was.

Instead, what’s been way more reliable is asking GPT to generate *phrasing*, not judgment. Things like:

“What’s a good way to write a Slack summary based on these numbers?”

“Reword this update into a format that’s less robotic and fits in one paragraph for a team newsletter.”

That version gives you bullet points like:
– “Customer satisfaction stayed steady this week with NPS at 54.”
– “Ticket response time continues to improve, now under 4 hours on average.”
– “Churn rose slightly, up by about 0.3 percentage points.”

It might not sound dramatic, but this helps massively with trust. I manually check and patch each GPT-generated summary, but most weeks I only change a sentence or two. The framing is there, the numbers stay intact, and the tone doesn’t swing wildly cheerful when things actually got worse.

Send the update somewhere visible but quiet

The destination of your weekly report matters more than you’d expect. My first few versions posted straight into the company-wide Slack channel as a thread. Big mistake. People misunderstood draft edits, reacted with emojis that implied growth when the numbers were down, and generally stressed me out 😂

So I moved it. It now lands in a private channel called #metrics-digest where only a few of us — heads of departments and product folks — hang out. From there, we read and forward it selectively. I also use Slack’s scheduled post feature to delay it until 10:00 AM, so even if the automation kicked off early, no one sees an incomplete version.

The Zapier step sends it like this:

– Checks the GPT output
– Inserts fallback text if GPT fails (just says “Report failed — check the Sheet”)
– Formats the Slack message with `*bolded headers*`, bullet points, and spacing

Every once in a while, if GPT totally chokes and gives me something bizarre (like a time it turned a customer NPS update into a haiku — yes, really), the Slack message just says:

> This week’s report is delayed while we sort out a formatting issue. Should be fixed soon.

Haven’t gotten yelled at yet 🙂

Make a quick button to rerun when needed

I can’t tell you how many times I ran the whole automation and then realized we were missing one number. Or someone tagged me at 9:12 saying “Oh actually, January refunds came in late, should we add that?”

Re-running the whole thing manually used to be a pain, until I made a Sheet button called ‘Rebuild Report’. This is linked to a simple Apps Script function that:

1. Re-cleans the sheet using the pre-made cleanup logic
2. Re-generates the summary block
3. Calls GPT via webhook again
4. Sends the output back to a review cell plus Slack

No need to open Zapier. No URL tweaks. Just a quick click.

BTW, the webhook sometimes gives a “403” error randomly when too many reports are triggered right after each other. I’ve added a rate limit pause of 3 seconds using `Utilities.sleep(3000)` — not elegant, but it seems to keep things happy.

Set up fallback log rows for debug sanity

There’s honestly nothing more soul-rotting than realizing your weekly report didn’t send, and you have no idea why. 😐 I’ve been there.

Now every single step sends a timestamp + result into a separate “Logs” sheet. Here’s what that sheet typically shows:

This helped me catch when the GPT webhook key silently expired (fun twist: the error was HTML, not JSON) and also when I accidentally passed an empty string as input. I highlight the failure rows bright red — just like my blood pressure when things break.

Try templating outputs instead of structuring inputs

My earliest approach had me painstakingly structuring the Google Sheets input like a JSON document. You know,:

“`json
{
“nps”: 54,
“churn”: 3.2,
“growth”: 1.4
}
“`

Sounds good in theory, but GPT doesn’t always follow key-based logic well. The better solution was to leave the Sheet in its semi-chaotic natural form and teach GPT how to generate into *fixed templates* like these:

> Weekly Update
>
> – Support tickets: X
> – NPS Score: Y
> – Churn Rate: Z
> – Revenue: $A
>
> Summary: [write here]

I provide about three message templates as part of the prompt and tell GPT: “If a number is missing, leave that line out.” Shockingly, that worked better than hand-coding every possible config.

My working theory is that GPT seems to respect structure more when you give it a template to fill in VS asking it to extract meaning from a messy blob. Moral of the story: teach the output, not the input, if you want sanity.