AI Guides › Playbooks
By Nigel Guy · 7 min read
Most people fight AI-sounding writing one draft at a time. They read a paragraph, wince at "delve" or "it's not just X, it's Y", type "make it sound more human", and get a fresh set of tics in the next draft. It feels like progress because each draft is slightly different. Nothing has been saved, so you repeat the correction every session.
The rule: have ChatGPT audit your own drafts for its tells, compress the findings into a short rule written as positive instructions, save that rule where ChatGPT reads it every time, then test the rule against a fresh draft. A rule you have not tested is a hope.
Four stages, each with a job and an output.
| Stage | What you do | What you end up with |
|---|---|---|
| 1. Collect | Gather 3 to 5 pieces of ChatGPT-drafted text you have actually rejected or heavily edited | A sample set, plus a note on what you changed |
| 2. Audit | Ask ChatGPT to find patterns across the samples, with quotes | A ranked list of tells, each with an example |
| 3. Compress | Ask it to turn the top patterns into a short instruction block that fits your character limit | A draft standing rule |
| 4. Test | Save it, generate a fresh draft on a new topic, then check it against the list | Keep, tighten or drop each line |
Use your own rejected drafts, not a generic list of "AI phrases" from the internet. The point is your tells in your context: the openers you cut, the closing summaries you deleted, the adjectives you swapped. Paste the originals, and where you can, your edited versions next to them. Your edits are the most useful evidence because they show what you wanted instead.
Fill in the square brackets and paste the prompt into a new chat.
You are a copy editor who specialises in spotting machine-written prose.
Context: I use ChatGPT to draft [TYPE_OF_WRITING, e.g. client emails, blog posts]. Below are [NUMBER] drafts it wrote for me, followed by my edited versions where I have them. Treat my edits as the evidence of what I dislike.
Goal: find the repeated habits in the drafts, the things I keep removing or rewriting. A good result is a short list of patterns, each tied to a quoted example from my text, ranked by how often it appears.
Steps:
1. Read all the samples before judging anything.
2. List patterns in these groups: openings, closings, sentence shapes, stock phrases, structure (lists, headings, summaries), and tone.
3. For each pattern give: a name, one quoted example, how many samples it appears in, and what my edit replaced it with (if I supplied one).
4. Rank by frequency. Drop anything that appears only once.
Constraints:
- Quote only text that is actually in my samples. Do not add typical AI phrases I have not shown you.
- If you are unsure whether something is a pattern or a one-off, say so.
- If I have not given enough samples to judge, tell me what else you need instead of guessing.
Before answering, check that every pattern has a real quote and a count.
Drafts: [PASTE_DRAFTS]
My edits: [PASTE_EDITED_VERSIONS_OR_WRITE_NONE]
Fill in: the writing type, the sample count, and your pasted drafts and edits.
The audit is a diagnosis, not a rule. Rules need to be short, because instruction space is limited.
You are helping me write standing instructions for ChatGPT.
Here is the pattern list from the audit you just did: [PASTE_RANKED_LIST].
Goal: write one instruction block of at most [CHARACTER_LIMIT] characters that makes future drafts avoid these patterns.
Requirements:
1. Cover the top [NUMBER_OF_PATTERNS] patterns only.
2. Write each line as what to do, not only what to avoid. For example, say how an opening should start, rather than only banning one opening.
3. Use plain sentences, no headings or decoration.
4. Add one line telling you to ask me before writing if the brief is missing facts, rather than padding.
5. Do not include any phrase you would then be quoting back at me as an example of what to avoid, unless a ban is essential. Keep any banned-word list to ten items or fewer.
Output: the instruction block, then its character count, then one sentence on what you left out and why.
Before answering, count the characters and cut until you are under the limit.
Fill in: your pattern list, your character limit (see below) and how many patterns to cover.
Save the block (next section), open a new chat, and ask for a draft on a topic that was not in your samples. Then run this check in the same chat.
You are an editor checking a draft against my standing instructions.
Instructions: [PASTE_YOUR_RULE]
Draft: [PASTE_THE_NEW_DRAFT]
For each line of the instructions, say whether the draft follows it, quoting the sentence that proves it. List any repeated habit in the draft that the instructions do not cover. Do not praise the draft. If you cannot tell whether a line was followed, say so.
Fill in: your rule and the new draft. Keep lines that held, rewrite lines the draft ignored, and add one line for each new habit that appeared.
Per OpenAI's help centre at time of writing, custom instructions are on all plans on web, desktop, iOS and Android. On web and desktop: Settings, then Personalization, make sure Enable customization is on, and paste into the Custom Instructions field. On iOS and Android the menu is Customize ChatGPT. Changes apply immediately to all chats, including existing ones.
The limit matters for stage 3: Free and Go plans can save up to 1,500 characters; Plus, Pro, Business, Enterprise and Education up to 5,000. Check your own plan page for current plan names and the £ price; this guide does not quote subscription prices because they change.
Two traps in where the rule lives:
Say you are a freelance consultant who uses ChatGPT for client update emails. You collect four rejected drafts. The audit finds: every email opens by thanking the reader for their message (4 of 4); every email closes with a sentence offering to "assist further" (4 of 4); and sentences keep following the "not only this, but also that" shape (3 of 4). The compress stage turns that into three positive lines: start with the answer in the first sentence; end on the next action and its date, with no offer of further help; make one plain claim per sentence. On a fresh email about a delayed invoice, the test shows the opening and closing held, but the "not only, but also" shape came back twice. You add a fourth line about that shape and test again. That is the loop working.