Set up the first prompt template even if it feels dumb
Let me tell you what I did the first time I tried using GPT to generate internal docs: I overthought everything. I built this whole fancy prompt chain for customer onboarding playbooks, pieced together seven different Zapier actions, and proudly hit “Test Trigger.” The doc content? Total nonsense. It was copying the structure but hallucinating teams that didn’t exist 🙂 Classic.
So, here’s the deal: when you’re building internal doc templates with GPT-style tools, don’t start with what *should* work. Start with the thing you wish someone had just typed into Slack last week. Like:
> “Can you summarize the last two sprints in bullet points, in normal human language?”
Write a plain-spoken prompt that does exactly that, no conditionals. Literally copy-paste how you’d ask someone:
“`
You are writing internal team updates. Summarize the Jira issues tagged ‘sprint-summary’ from the last 14 days.
Use plain language, short bullet points. Avoid duplicate info. If a ticket has no clear resolution, skip it.
“`
Whenever I try to get fancy — stack conditions or include instructions like **”do not invent information”** — GPT almost *immediately* ignores it, especially if the context window’s full. You’ll get better results by:
– Writing out example inputs and expected outputs
– Adding simple logic rules after the main request
– And framing your tone *inside* the prompt with words like “This is casual. Imagine you’re explaining this to the new team member.”
Every time I tried to reference “company values” in the prompt, it invented new ones.
Test your prompt on bad and weird inputs
If your input source includes inconsistent ticket titles, half-filled forms, or worse, free-text Slack inputs — congrats, you need to deal with garbage data 😅
Your prompt template can’t just be for the happy path. I literally had GPT confidently generate doc sections about tasks that had no description and weren’t even assigned to a real person. My favorite was the one that listed:
> **Owner:** Unknown entity – likely responsible
Cool. So.
Here’s how I test:
1. Pull some failed or null entries from your trigger (mine was a Jotform logic branch where people didn’t finish the form). Feed those into your prompt as-is.
2. Drop in examples where the internal tags don’t match — like if the status is marked “done” but the notes field says “please revisit.”
3. Give it overlapping or nearly duplicate values, like “feature request” vs. “feature idea.”
Then I modify my prompt slowly over 3–4 test cases, not all at once. I’ll write things like:
“`
If the source entry has missing tags or notes, write: ‘No useful content provided.’ Never invent.
If the title and description seem to repeat, only include one.
“`
That last line saved me so many hallucinated paragraphs.
You’ll get to a point where your GPT-generated internal doc actually says useful things like:
> – Added support for dark mode in Settings panel (ticket 82)
> – Timeline update postponed, no issue summary provided
> – Removed deprecated endpoints – skip
That’s when I trust it in automation.
Save the base prompt somewhere version controlled
This is so dumb but I didn’t do it until I broke everything.
I adjusted some suffix in the prompt template — just rearranged the sentence that said “skip items without title or assignee”— and suddenly it stopped omitting the bad entries. I kept re-testing, thinking OpenAI was being flaky (it was not this time), and after two coffees, realized I had removed the phrase “never include empty results.” Sigh ¯\_(ツ)_/¯
I save every working version of a prompt template in a Notion DB and give them weirdly specific names:
– **🟢 sprint_summary_clean_v3_mdstyle**
– **🟡 partner_intake_dirty_keywords_inline**
– **🔴 form_rescue_attempt_2_should_be_deleted**
This lets me diff old vs. new prompts when the output quality mysteriously tanks. Also? I link directly to where they’re used: like if it’s plugged into a native Zapier OpenAI step (not Code), I’ll label that. If it’s inside a Make webhook, I sometimes forget whether I’m feeding in markdown or raw text, and headers break everything.
Bonus note: GPT sometimes quietly chokes if your prompt has a weird invisible character. I pasted a curl command from a doc that had a non-breaking space in it (thanks, Google Docs), and it caused the entire prompt to misalign when pasted into Airtable’s script automation.
Real formatting issues that broke my internal doc bot
Formats matter. Here’s what broke my Zap-template combo:
– GPT inserting smart quotes (“ ”) instead of straight quotes — especially bad in YAML snippets
– URLs getting rewritten with extra trailing slashes (which broke preview cards)
– Extra line breaks between bullets when copied into Notion
One time it started capitalizing every internal tool name like this:
> Used Jira, Linear, and Monday.Com to reconcile sprint scope
No one on my team says “Monday.Com.” It looked obviously AI-written.
I now include explicit rules in prompts like:
“`
Spell internal tool names exactly as in the list below: Jira, Linear, Notion, Slack, Zoom, Figma.
Do not capitalize domain-level terms (e.g. Monday.com → use Monday)
“`
Also, sometimes GPT would change internal acronyms. Instead of writing “PTO,” it rewrote it as “vacation time” in one case, but left “PTO” in others. That inconsistency made reviewers think it was wrong. It’s worth freezing terms like that.
Let your team break it a bit before trusting the output
I always want to ship these bots day one. But then someone tags a Jira issue with six labels and an emoji in the description, and the doc format goes Feral.
I now send one of these messages in a channel before turning on auto-publish:
> “Hey I’m testing a little summarizer bot — it’s pulling from #intake-requests if the label ‘doc-me’ is added. If your request looks funky, ping me. It’s probably broken and that’s fine.”
In 90% of cases, someone’ll paste in a weird edge case I never thought of. Like someone added an image to a Linear comment and the summary said:
> Referenced media – possibly Figma assets or memes
Idk. That line was not helpful.
So I pause, update the prompt — like “Ignore comments with only images if there’s no caption text” — and rerun things until the output gets boringly reliable. GPT is excellent at being *approximately right* and *confidently off*. Embrace that.
Add human phrasing back into the prompt regularly
After heavy refinement, prompts tend to get stiff. If you have too many conditionals and phrases like “If title is empty OR status is unknown AND category is flagged: skip,” the bot starts sounding like it’s filling out a tax form.
I add little phrases like:
“`
You’re writing this for a busy teammate who only has 2 minutes. Be brief but natural, like you’re writing a Slack message.
“`
That’s usually enough to inject human rhythm into the bot’s flow. You can always fix spacing or format later. But a doc that keeps the reader awake is way more valuable than one that checks every edge case but is unreadable.
GPT’s tone drifts every few weeks based on model updates. Periodically re-run your final template with fresh inputs — sometimes it starts sounding robotic again for no reason and needs recalibration.
The trick is being okay with rewriting your own fix. Often.