Starting with messy draft prompts
If you have ever tried to use GPT4 to write emails for you, then you probably know the first thing you get back always feels a little too clean. When I first opened ChatGPT and typed something like “write me a professional email to a client about rescheduling a meeting,” the result looked okay at first glance. It had the right words, the right shape, but when I copied it into Gmail it felt so stiff. The client would know instantly that this wasn’t written in my normal voice. Even the greeting sounded robotic, and the closing line read like a template I had stolen from some HR handbook. That is when I realized I had to tweak the actual prompt every single time instead of just copy pasting. The frustrating part was I never wrote prompts the same way twice so I had half finished drafts scattered all over Notion, Google Docs, and Zapier test steps. And of course some of these worked once, then suddenly refused to work again without me knowing why.
Figuring out what GPT4 misunderstood
The biggest problem I had was tone. For example, I would ask it to write a “friendly but professional” email. That phrase sounds obvious in my head, but GPT4 interpreted it like a corporate HR training module. Words like “synergy” or “aligning schedules” crept in, and I cringed reading them. I had to experiment with describing tone in much plainer ways, like “write it like a coworker who is polite but not formal.” That tiny change instantly made emails sound less like a robot. If you are new to this, a trick that helped me was pasting one of my real old emails at the bottom of the prompt and just saying “match this style.” The model suddenly imitated little quirks, like the way I tend to put “Sounds good to me :)” as a closer. That felt much closer to my real voice. It’s not perfect every time, but you start to notice which awkward phrases keep showing up and then you can throw in extra instructions like “avoid saying synergy.”
Adding context directly into the prompt
One of my biggest mistakes was assuming GPT4 would just remember context. When writing followup notes to a client, I would think “it knows the last email thread because I pasted it earlier.” But then I’d copy a shorter version of the prompt into another tab and wonder why GPT4 started writing a cold pitch email instead of a friendly followup. That happened so many times. I finally got into the habit of putting a quick table above my draft so GPT4 could reference it. For example:
Client Name | Purpose | Notes
———– | ——- | —–
Sarah | Meeting reschedule | She is usually casual and signs off with just her name
This made a huge difference because GPT4 worked with the structure right above the email instead of inventing details. If you accidentally forget the notes column, the output tends to drift back toward generic filler like “I hope this email finds you well” ¯\_(ツ)_/¯.
Creating reusable prompt templates
After too many half broken experiments, I finally tried to make reusable templates. Nothing fancy — just a Google Doc with different blocks I could paste into ChatGPT. The problem though was that every template needed little tweaks. For a job application followup, my template said “thank you for your time” but once I pasted it for a casual vendor email, it sounded way too stiff. That is when I started marking optional lines with brackets like [insert casual closer here]. By treating templates as half finished, I avoided the trap of expecting them to be one click ready. The funniest part is I had built Zapier automations trying to generate whole emails without me looking, and they always ended up unusable because the tone was weird. The brute force version where I still do the last 10 percent edit is actually more reliable.
Common errors that confused me
Sometimes GPT4 output would just cut off weirdly. I would get a great opening line, then the body would be two sentences and stop mid thought. At first I thought I broke something, but it was just the length limit with my free tier at the time. Another confusing bug came when I asked for super short responses like “write this in three sentences.” GPT4 somehow took that as “be vague” and the content sounded empty. What actually worked better was telling it the specific things to include, like “start with acknowledging their concern, share one quick detail about the project, and end by proposing next week.” The output stayed short but no longer hollow. One time I tried using Zapier to feed Slack messages as prompts, and for no reason the Zap fired twice in a row and sent two similar but mismatched emails into my drafts. I almost hit send before I noticed. Ever since, I always test Zapier automations by pointing them at a dummy inbox rather than my real one. 😛
Improving results with roleplay instructions
This might sound silly, but telling GPT4 to roleplay helped a lot. If I just said “draft a polite email,” the response was still flat. But when I wrote “pretend you are me, someone who is slightly informal but reliable, and you need to convince a client kindly without sounding desperate,” the emails looked like something I actually would have written on a late night. Adding that “pretend you are me” line mattered more than I expected. I even experimented with “pretend you are my coworker writing as me” just to shake the structure loose. Sometimes that added a strange phrasing I would never use, but it still gave me fresh text I could steal snippets from. The point is, don’t just give instructions. Anchor them to how a real human persona would think when writing the email.
Checking output before pasting it
I cannot count how many times I copied an email into Gmail without reading it carefully enough. Once it literally had the phrase “INSERT NAME HERE” left in the closing, because I had written the placeholder in the prompt. Another time GPT4 added a strange motivational sentence at the end like “Together, we will achieve great results” — my client would definitely laugh at that. Now I have a small rule. I paste the draft into Google Docs first, highlight any parts that look suspiciously smooth, and replace them with my normal phrases. Think of it like sanding a piece of wood down, the edges are sharp at first, then you smooth it just enough but not so polished it looks fake. I even started keeping a list of phrases GPT4 loves to use (“moving forward,” “I hope this finds you well”) and I swap them out before hitting send.
Making it work with automation tools
Eventually I tied all of this back into automation because copying and pasting every single email draft was too slow. My setup now is a mix of Zapier and Make that pushes calendar updates into a Google Sheet, then sends the context table plus instructions into GPT4. The draft comes back into a “Drafts” folder that I still have to approve. That last step is essential. I wasted a week trying to get rid of manual approval, dreaming of totally hands off email drafting, but the output just never passed the vibe check. Someone always caught one line that felt slightly off. The automation works best when it gives me a head start instead of pretending to take over the whole job. If you are curious, tools like openai.com make it easy to connect API access without too much coding, but be ready to babysit it in the beginning. Once you catch the weird glitches early, the workflow calms down.
Small wins that felt like breakthroughs
The first time I got GPT4 to mimic my casual signoff perfectly felt ridiculous but satisfying. It ended with “Talk soon, -Josh” which looked exactly like me, not like a template. Another small win was when I figured out I could say “keep the sentences under 12 words each,” and the flow immediately sounded more human, less padded. The funniest was when GPT4 once apologized inside the draft like it was itself writing a correction — I actually kept that little humanlike slip because it felt believable. These tiny wins made me trust the system enough to keep building on it, even after half the earlier experiments collapsed for no clear reason. Every once in a while I still see the old stiff tone bleed back in, but now I know exactly which part of the prompt probably caused it and how to patch it next time.