Custom Instructions for ChatGPT That Fix Your Email Drafts

When you ask ChatGPT to draft a professional email without custom instructions, it has no signal about who you are or how you write. It reaches for the safest option in its training data: formal corporate register, hedging language, and phrases that appear in millions of business emails. The draft technically answers your request. It sounds like no one in particular wrote it.

Custom instructions are the persistent settings that load before every ChatGPT session. Open Settings, then Personalization, then Custom Instructions, and whatever you put there runs as context on every conversation. The problem is that most people write instructions too vague to change anything. 'Write in a professional tone' does not help, because the model already considers its defaults professional. What it needs is your role, your register, a sentence-length ceiling, a list of the exact phrases you avoid, and a rule for what to do when you leave out a detail.

This guide walks through each element with copy-paste blocks for three common contexts: founder, support, and client-facing email. The final section covers the one thing custom instructions cannot hold — and why that matters once your drafts depend on relationship context, not just writing preferences.

What 'sounds generic' actually means in a ChatGPT email draft#

A draft sounds generic when it could have come from anyone, to anyone, about anything broadly similar. There are three failure modes, and the fix is different for each.

The first is wrong register. ChatGPT defaults to formal-corporate when given no signal about how you write. 'Please do not hesitate to contact us should you require further assistance' is the output when no register is specified. The fix is to name your actual register explicitly in the instructions.

The second is invented context. You write 'draft a reply to the vendor about the delayed delivery' without naming the vendor, the original date, or the tone of the relationship. The model fills the gap with plausible-sounding specifics. You spend more time removing made-up details than you saved by using AI. The fix is a missing-facts rule that tells the model to ask before guessing.

The third is recurring banned phrases. 'I hope this email finds you well,' 'touching base,' 'circle back at your earliest convenience' — these appear because they are heavily represented in business-email training data. Telling the model to 'avoid filler' does not remove specific phrases; the model's judgment of what counts as filler differs from yours. The fix is an exact list.

Fix 1: Set your role, register, and sentence-length ceiling#

This block goes in the 'What would you like ChatGPT to know about you?' field. It tells the model who is writing, what tone that person uses, and a hard ceiling on sentence length. These three lines do most of the corrective work.

Without a role, ChatGPT has no way to choose the right register. A founder writing to an investor, a support agent replying to a ticket, and a salesperson following up on a proposal all need different tones. The model cannot pick the right one without a signal.

1. 1

   Open Settings and navigate to Custom Instructions

   Click your account icon, choose Settings, then Personalization, then Custom Instructions. The setting is available on free and paid ChatGPT accounts. If you are inside a custom GPT or a workspace account with an organization prompt, those prompts override your personal custom instructions — return to a standard chat to test your changes.

2. 2

   Paste your role, register, and sentence ceiling into the first field

   In 'What would you like ChatGPT to know about you?', replace any existing text with a block that names your role, your register, and a word-per-sentence limit. The example below is a starting point. Edit it to match how your emails actually sound — the more specific you are, the more the output shifts.

3. 3

   Test with a short draft and check three things

   Ask ChatGPT to draft a two-sentence reply to a hypothetical colleague. Check the opener (did it skip the standard greeting?), the sentence length (are they inside your ceiling?), and the register (does it sound like you?). If any of the three miss, make that line in your instructions more specific — add one example of what sounds right and one of what does not.

Fix 2: Add the banned-phrases list and the missing-facts rule#

Fix 1 sets the positive instruction — how to write. Fix 2 sets the negative constraints — what to avoid and how to handle missing information. Both go in the same 'What would you like ChatGPT to know?' field, appended below the role block.

The banned-phrases list works better than a tone instruction because it is specific. 'Avoid corporate language' is a judgment call the model makes poorly. 'Never write I hope this email finds you well' removes a particular phrase and leaves no room for interpretation.

1. 1

   Append the banned-phrases section below your role block

   In the same custom instructions field, add a line break and a section labelled 'Phrases I never use in email:' followed by the specific phrases you want removed. Start with the list in the example below, then add any that keep reappearing in your own drafts despite the instruction.

2. 2

   Add the missing-facts rule directly below the banned phrases

   The rule is one sentence: if ChatGPT needs a name, date, number, or relationship detail that you have not provided, it asks before drafting. Without this rule the model will invent plausible-sounding specifics every time it encounters a gap, and edited-out invented facts take longer to clean up than just supplying the detail in the first place.

3. 3

   Test with a deliberately incomplete prompt

   Save the updated instructions. Ask ChatGPT to 'draft an email to the team about the project delay' without naming the project, the date, or the reason. With the rule in place, it should ask for those details before writing. If it guesses instead, restate the rule more directly: 'Never infer missing facts. Always stop and ask me first.'

Fix 3: Choose a role variant for your specific email context#

A general role block covers most email. Three contexts need a tighter variant because the expected register is meaningfully different: founder email (direct, economical, peer-to-peer), support email (warm, solution-first, no blame language), and client-facing professional email (formal enough to be taken seriously, specific enough to sound human).

Changing the variant means changing one line — the register line. The banned phrases and the missing-facts rule stay identical. Swap the register line for one of the three below depending on your primary email context. If you write mainly in one mode, make it the permanent default.

1. 1

   Identify which context covers most of your email volume

   Founder emails are direct and peer-to-peer: they skip pleasantries and lead with the point. Support emails are warm and solution-first: they acknowledge the issue before the resolution and avoid defensive language. Client-facing professional emails are formal enough for the business relationship but specific enough to sound like one person wrote them, not a department. Pick the one that fits the majority of what you send.

2. 2

   Replace your register line with the variant line

   In your custom instructions, find the sentence that describes your current register. Replace it entirely with the variant line from the example below. Keep everything else — the role description, sentence ceiling, banned phrases, and missing-facts rule — unchanged.

3. 3

   Run a test draft and verify the shift

   Ask for a typical email in that context. The opener, sentence structure, and level of formality should all reflect the new register. If the shift is not obvious, add a concrete contrast: 'Sounds like: Hey Sarah, two things.' and 'Does not sound like: I am writing to bring to your attention the following matter.' Concrete examples are more reliable than abstract description.

How to tell which fix you are missing#

You do not need all three fixes simultaneously if one problem is clearly dominant. Read your most recent AI-generated email draft and look for these symptoms.

If the opener is 'I hope this email finds you well' or 'I wanted to reach out to touch base' — that is the banned-phrases problem. Fix 2 solves it directly. If the draft reads like a press release — passive voice, overly formal, every sentence hedged — that is the wrong-register problem. Fix 1 solves it. If the draft contains a date, name, or detail you did not provide, and that detail is either wrong or invented — that is the missing-facts problem. The rule added in Fix 2 stops it.

If all three symptoms appear together, implement all three fixes in order and test between each step. Most people find that Fix 1 and Fix 2 together account for nearly all of the generic quality. The role variant in Fix 3 matters most once the base block is working and you notice the tone still shifting in ways the single role description does not handle.

Preventing it: the full copy-paste block#

The block below combines all four elements into a single paste. Swap in your role, choose a register variant, add any banned phrases specific to your writing, and save. From that point, the instructions run automatically on every new conversation without re-entering any of this context.

Custom instructions have one structural limit: they hold your preferences but not your relationships. They know your register, not which contacts get the direct founder voice and which get the more formal client-facing one. They know you avoid 'touching base,' not that the last email in this thread ended on a difficult note and the next one needs to acknowledge it. A chat window with standing instructions knows how you write; it does not know who you are writing to or what the history is. AI Emaily handles that differently — it holds a user-set Personal Context brain and per-client profiles you configure, so drafts are calibrated to both your voice and your specific contact, with that context drawn from your actual inbox at draft time rather than re-entered each session. We build AI Emaily. You can read how the drafting works at aiemaily.com, and a 7-day trial at aiemaily.com/pricing lets you test it against your real inbox before committing.