AI Prompts & Workflows Archives - Page 13 of 33

BuildYourPromptTemplateBeforeYouDoAnythingElse

If you’re trying to generate product descriptions in bulk with GPT, the first thing you absolutely need to do is lock down your prompt template. I made the mistake of skipping this. Twice. Both times, I ended up with 200 descriptions that all started with some version of “This product offers great features.” Awful 😑

Here’s a simple way I approach it now.

1. I grab a few actual samples — not what I *want* the description to sound like, but what decent ones on the site already look like.
2. Then I write a single prompt in plain English explaining what GPT needs to do, using brackets to mark replaceable parts. For example:

“Write a short, engaging product description for a [product_type] called [product_name] that has [key_features]. Target audience is [audience_type]. Include 1 power benefit and a metaphor.”

Same prompt will work across a JSON, CSV, or even a giant Google Sheet.

If your prompt is too vague (like “Generate a good description”), you’ll get sections like:
> “This item will help you in many ways.”

Nope. Be specific. Oh and test it with a few weird edge case products — the ones with long names or oddly specific ingredients.

SetUpYourGoogleSheetOrCSVWithVariables

Now comes the part where you need to prep your structured data. This becomes your merge source later — think mail merge, but for AI.

You need at least 4-6 columns, depending how detailed you want the descriptions. Mine usually look like this:

Keep the formatting clean. No trailing spaces. And whatever you do, don’t leave cells randomly blank. GPT will just hallucinate stuff. I once gave it a CSV with missing “audience_type” and got a paragraph about motorcycle repair enthusiasts trying to buy yoga blocks. 😐

Google Sheets is nice here because you can use formulas and cleanup tools — like TRIM() to strip whitespace. If you’re using Airtable or Notion, just make sure you’ve got clean exports.

ChooseYourPromptRunnerAutomerTools

You can pipe this data + prompts into GPT using a few different tools now. Here are the ones I’ve tried:

1. **Airtable GPT Automations** — cute idea but it throttled like crazy once I ran the bulk. I’d come back 20 min later and only 4 descriptions had generated 😫

2. **Zapier + OpenAI action** — solid but pricey if you’re generating hundreds of items. Also doesn’t show previous GPT history in any visible way when debugging failed runs.

3. **Make.com OpenAI module** — honestly pretty flexible. You can plug in a table and loop each row with a Data Iterator, using Output Mapping to grab columns from Google Sheets.

4. **GPT for Sheets and Docs** — my current favorite for fast prototyping. You just use `=GPT_FILL(prompt_template, variable1, variable2…)` right in the cell.

I’ve had the best balance of cost and speed using GPT for Sheets plugin with a batch run script. You can run it in blocks of 50 descriptions, add a 3-second pause between rows, and avoid API errors.

SetUpRateLimitsAndBatchingToAvoidCrashes

Here’s the nasty bug I hit: after a few hundred rows, GPT just started returning errors randomly. No helpful message. Sometimes it would just output the word “null.” Other times, nothing at all. I reported it; GPT said it was a timeout. So I set up batching with waits.

In Google Sheets, I split the prompts into groups of 25 using this in a helper column:

`=ROUNDUP(ROW()/25, 0)`

Then I filtered and ran each batch manually (ugh, I know), but it actually saved time compared to debugging broken output rows afterwards.

If you’re using Make or Zapier, use a delay in your Loop module — I had best results around 1.8 seconds between iterations. Go faster and it’ll silently fail at totally random points. Not a fun way to spend your afternoon 🥴

AddCustomFlavorWithConditionsInYourPrompt

Once your pipeline is generating text, you’ll probably want to include things like tone of voice, languages, or variations.

Instead of making separate prompts for each tone or vibe, I do it with mini conditions like:

“If the brand_tone is ‘fun’, use light humor and playful phrasing. If it’s ‘premium’, keep it elegant and minimal.”

It works surprisingly well — though sometimes it overdoes it. One time the “fun” tone gave me:
> “This gadget’s got more tricks than a magician’s sleeve.”

…which, okay, calm down, GPT 😅

Other variables that are fun to play with:
– Description length option (short vs. medium)
– Add a testimonial-sounding sentence
– Include SEO keyword if present

You can pass these all from your Sheet as more columns.

PostProcessTheDescriptionsBeforeGoingLive

Even with perfect prompts, you’ll probably find stuff you need to fix. I now use the following post-processing steps:

1. Run a quick visual scan down the descriptions column. Check for weird line breaks or repeated phrases. GPT loves repeating lines like “You won’t be disappointed.”

2. Search for banned phrases. I use this formula to scan for things like “amazing quality” or “must-have”:

`=ARRAYFORMULA(IF(REGEXMATCH(A2:A, “amazing quality|must-have|best ever”), “⚠️ Fix”, “”))`

3. Fix brand name inconsistencies. GPT will sometimes call something “Our XYZ” when the company never uses that phrasing. I use CTRL+F to scan for our/we/us in the cells.

4. Capitalization check. Look especially at model names or size abbreviations—it’ll randomly lowercase proper names if you didn’t include them in the prompt.

Most annoying bug I ran into was GPT pasting actual quotation marks inside made-up testimonials I never asked for. Had to bulk remove quotation marks using Google Sheets formulas. Not proud. But it worked.

SaveTheFinalCopiesSomewhereSafeBeforeYouTweak

Before you start rewriting or A B testing, save the raw outputs somewhere untouched. I’ve made the mistake of overwriting my original AI drafts while adjusting tone, and then couldn’t tell which pieces were human-edited vs. generated. 😭

Solution: Drop all original descriptions into a separate tab first. Label the column “raw_export.” Then create a second tab for “edited_description.” That way, if you break your rewritten version halfway through, you can always go back and reference the AI’s original idea.

Also helps especially when… let’s say a stakeholder decides halfway through that they, quote, “don’t want the product page to sound so AIish.” Cool cool. Let me just redo 312 items by Thursday.

MonitorForEmptyOrRepeatingResultsAfterChanges

Here’s something that once nuked half my batch: I made a small tweak to the prompt to improve the call to action. What I didn’t realize is that in 50 rows, GPT just… gave up. It started repeating the same closer: “Get yours today before it’s gone!”

I didn’t catch it until a coworker said the page “felt like an infomercial.” 😒

Now I run a duplicate-phrase checker every time:
– Copy the descriptions column
– Paste into a free word frequency tool
– Look for high counts on phrases like “perfect for” or “don’t miss”

Also: if GPT ever starts returning BLANKS mid-batch, it’s often an issue with your prompt not handling a certain edge case (e.g., missing feature, empty name, etc). Add some If logic to handle nulls.

“If [feature] is present, explain its benefit. Otherwise, skip this sentence.”

That little snippet will save your sanity.