Using AI to Prompt Daily Standup Check-Ins Without Confusing Your Team

Using AI to Prompt Daily Standup Check-Ins Without Confusing Your Team

ByUS Stacked

1. Building a simple check-in flow inside Slack with Zapier

I used to just ping our team in Slack every morning with a shared calendar event that said “Daily Standup” and hoped people remembered what to post. Not surprisingly, it became a dumping ground for unrelated updates or, worse, silence. So I rebuilt it using Zapier + OpenAI + Slack’s workflow builder. The idea was simple: create context-aware prompts that nudge people with specific check-in questions based on recent work.

The Zap is triggered at 9:00 AM on weekdays. It searches each teammate’s recent commits from GitHub (or comments from Asana, depending on the team) via an RSS parser I set up in Feedly, then passes that to a GPT-4 prompt that spits out a personalized message like “Hey Julia, I noticed you merged the onboarding flow last night—any blockers before launch?”

That part works. What doesn’t? Slack’s native threading behaves weirdly when the bot posts via Zapier. Sometimes replies end up in a fresh thread instead of under the original prompt. There’s no consistent message_ts behavior when OpenAI appends links inside the body text—it forces a preview, which offsets Slack’s internal layout accounting.

Half the team didn’t notice the AI-crafted message was even talking to them until I bolded their first names with markdown. Then it clicked. That tiny formatting shift had a bigger engagement impact than the personalization logic.

2. Using OpenAI to generate questions from past activity logs

I didn’t want to write a new prompt format from scratch every time we added a new data source. I made the mistake of reusing my Airtable-driven prompt from a client project thinking it’d plug in identically. It didn’t. The structure of activities in each tool matters: Notion exports have embedded timestamps in a totally different position than Asana’s API, which clusters updates under a “page_updates” object that wasn’t documented anywhere.

The prompt had to be updated to anticipate nested context. Here’s an early version that broke because the JSON formatting exceeded the OpenAI 16k token limit without warning:

{"activities": [...], "user": "Julia", "format": "Check-in question"}

After trimming and summarizing the activities in a separate step using GPT before the final prompt, I clocked that the model stopped hallucinating people’s names. Apparently, when you feed the model more than one activity related to someone, it starts making up contributors who don’t exist in your org. That behavior didn’t show up on the playground, only in production via Zapier’s OpenAI actions. Reproducible? Only sometimes. My guess: token noise from improperly structured feeds.

3. When teammates ignore your prompt because it looks like noise

Let’s be honest, almost every automated Slack message eventually looks like background radiation. The first few days of prompting worked great—real responses, follow-up threads, even emojis. But by day five, engagement cratered. I thought the prompt quality was degrading. Turns out, it was the delivery method itself.

Messages were posted as Zapier Bot instead of my actual name. The human layer was gone. People subconsciously tuned it out. I switched the message sender to my own Slack identity via a webhook URL posted with a Slack token, and instantly, replies came back.

Also, unless your prompt ends with a direct question mark and two lines or fewer, people treat it like an announcement. I tested one version that ended in “…anything in the way?” vs “Let me know if anything seems blocked today.” The latter got zero responses multiple times.

If your AI message reads like an FYI update, people act like it’s optional. Even if it’s asking them something.

4. Slack threading and Zapier message updates do not mix well

Tried updating a message after it posted to fix a broken link once. Zapier’s Slack action says you can pass message_ts to update an existing message. But in multi-step Zap runs, especially when branching paths determine which teammates to ping, Zapier loses the message_ts because each thread launches in a separate context.

The hack I landed on was capturing the response from the message posting action in a named variable, storing it temporarily in Storage by Zapier, and then calling it from the next step. But that only works if no parallel paths are active (like if you fork responses to multiple users in the same Zap). Otherwise, race conditions overwrite stored values before you finish reading them.

Also worth noting: if the Slack workspace has custom emoji or disabled rich previews, GPT-generated text behaves erratically. It sometimes cuts off mid-sentence if the message includes a text entity that Slack blocks—especially when special characters are auto-rendered.

5. Prompt tuning experiments that actually improved response quality

I burnt half a Saturday rewording the AI prompt and got nowhere until I realized: you can’t just ask for “a personalized daily question”—you have to define what makes it interesting. The breakthrough came when I added a few examples of engaging vs boring check-ins right in the system message. Like this:

System:
Good: What did you learn yesterday building the pricing toggle?
Bad: What are you working on?

That immediately increased GPT’s specificity. Without it, the AI would default to stale patterns like “Any blockers today?” or “Do you need help?”. Now it reliably ties to actual artifacts—pull requests, recent edits, page views.

Also: setting temperature to 0.5 avoided those weird overly casual “Hey buddy!” messages that creeped some folks out. GPT loves making jokes if you don’t tell it not to.

  • Use OpenAI’s system message format and show examples inline
  • Drop the temperature slightly to suppress unnecessary flair
  • Strip emojis unless GPT can see emoji usage in prior messages
  • Avoid “daily questions” framing, use “Reflect on…” instead
  • Limit prompt length—verbose inputs reduce action relevance
  • Include teammate names in the Assistant’s role definition
  • Test outputs in your actual Slack emoji/font theme context

6. Edge cases where timezone logic broke everything silently

At first I let the Zaps just run at 9 AM in my local timezone, forgetting that we had distributed contributors. One guy in Romania kept missing his check-in because it landed at 5 PM for him—post standup. Once he replied “This prompt’s always late” and I had to dive into the setup.

Turns out Zapier’s “Schedule by Zapier” trigger doesn’t account for user timezones unless you specifically set a static UTC offset, and even then, Slack’s display timestamp adjusts separately for each user. So on his screen, it always looked late—even if it wasn’t.

The fix was rewriting the flow to send separate prompts in staggered blocks using a lookup table based on user email → country. Not elegant, but reliable. I stored city-by-city offsets in Airtable and matched them via OpenAI to infer context language.

There’s another fun bug: Slack only shows one scheduled message indicator per user if sent within a one-minute window—even if you’re firing ten Zaps. So it looks like only one message posted, while others silently arrive. Multiple people missed prompts entirely until we spaced them 60 seconds apart.

7. What happens when teammates copy your system prompt without reading

I shared the entire Zap with a teammate using Zapier’s “Share a Zap” feature. They set it up for their team, but dropped in their own OpenAI key and changed the Slack channels. I got a DM two weeks later: “This thing’s yelling at Sean about files he didn’t touch.”

I looked at their Zap history—turns out, they rewired the activity input feed to pull from comments in ClickUp, which includes assigned users and their supervisors. GPT was pulling random references to mid-level managers and attributing actions incorrectly because the context didn’t filter by user. The original Zap had that logic baked into a formatter step.

This showed me: prompts aren’t portable. Your team’s data shape, tone, tooling—all of it affects how GPT behaves. And if a manager thinks AI is calling them out for something they didn’t do, they stop responding completely.