To get a usable two-sentence apology to a landlord, specify the length, audience, facts to include, and what to leave out. Add a sample of the tone you want. Otherwise a chat may return 400 words, a subject line, and a legal lecture.
It is 10:41 p.m. The notice is on the left and ChatGPT is on the right. You type “write a polite email about the late rent” rather than stare at a blank Gmail draft. The reply reads like a cover letter. You spend the next ten minutes deleting paragraphs you never asked for. The prompt left too much room for the model to fill.
Coding assistants have a similar problem. A Korean walkthrough of Claude Code puts the first pass in a plan step, lets you edit the plan, and only then writes. For an email, define what a finished reply looks like before asking for the draft.
Why does ChatGPT ignore half of my prompt?
The two-sentence limit can disappear under six other jobs. Models may focus on the last clear task they can see. Stack “be professional,” “be warm,” “don’t sound like AI,” and “keep it short,” and the length limit may lose out. Put the length and audience in the first two lines.
Anthropic’s prompting notes compare the model to a new hire: “Think of Claude as a brilliant but new employee who lacks context on your norms and workflows.” Give that employee four jobs without a priority order, and “write a professional email” can crowd out “two sentences.”
Consider this overstuffed Sunday prompt:
Write a polite, professional, friendly, firm, not-too-formal email to my landlord about paying rent three days late. Don’t sound like AI. Don’t be legal. Don’t be groveling. Keep it short but cover the reason, the date I’ll pay, and an apology. Make it good.
The resulting 280 words covered the reason but also invented a hardship story, offered to set up autopay, and closed with “please don’t hesitate to reach out.” None of that was in the lease. The prompt specified plenty to include and little to leave out.
A shorter prompt sets a clearer boundary:
Task: two-sentence email I can paste into Gmail. Audience: landlord who already emailed once. Facts: rent is three days late; I pay Friday by Zelle; no extra story. Done when: I can send it without deleting a paragraph. If a fact is missing, ask one question instead of inventing it.
The last line tells the model what to do when information is missing. A coding lecture treats “done” as four checks: the happy path, the error message, a save you can see, and whether the next step still works. The email equivalent specifies a good reply, a response to missing facts, and an output format. Otherwise the model may fill the gaps with plausible fiction.
Conflicting tone words are worse than missing ones. “Friendly” and “firm” can live together. “Short” and “cover seven topics” cannot. Pick a word count or a sentence count. Put it above the rest.
Should I paste an example or just describe the tone?
Paste an example. “Casual but professional” leaves more room for interpretation than an email you sent last month. Google’s Gemini prompting guide says: “We recommend to always include few-shot examples in your prompts. Prompts without few-shot examples are likely to be less effective.”
You do not need five. One or two is enough for a landlord note, a meeting recap, or a product listing. Three starts to help when the job is a format the model keeps missing, like a table with four columns or a JSON blob with exact keys.
The example has to be the output, not another lecture about the output. This fails:
Write in a warm tone like a good coworker.
This works:
Match this email I sent in March: “Hi Dana — the March rent hits your Zelle on Friday. Sorry for the delay.” Same length. Same leftover chill. New facts: April, three days late, Friday.
Google’s own docs show the same trick with explanations of snow. Ask “which explanation is better” with no samples, and the model picks the long one. Show two short answers labeled as the winner, and it picks the short one on the next question. The model copies the pattern you rewarded.
A style guide can grow as long as the lease: “Never use the word utilize. Never start with ‘I hope this email finds you.’ Never use contractions. Always use contractions in the greeting.” Those conflicting bans compete with the example. A review of switching coding tools makes a similar point: name the constraints you need and drop the rest. Do not spend a sentence banning “utilize” unless that word has caused a problem.
If you cannot find a past email, write the first sentence yourself and ask the model to finish in the same voice. Half a sample beats a mood board.
How much of my email thread should I paste?
Paste the latest landlord message and the three facts you want in the reply. Not the six-month thread. Extra pages do not make the model more careful. They give it more material to quote, summarize, or argue with. Put the long text first, then your question, if the chat tool does not already pin the file.
A full thread may look thorough, but the model can answer the oldest complaint or write an unwanted timeline. In a contractor-dispute example, fourteen pasted messages led to a draft that apologized for a leak already fixed in March. The archive contained that fact; the problem was which message needed a reply.
A working paste looks like this:
Source, latest email only: “Please confirm the Friday payment and the late fee.” My facts: I pay Friday. I am asking the late fee to be waived once. I will not argue history. Output: five sentences max, no bullet list.
If you have a PDF, say which pages matter. “See the attached lease” is not a page number. “Section 4, late fee, first paragraph” is. When the file is long, ask the model to quote the line it is using before it writes the reply. That is slower. It also stops it from citing a clause that is not there.
People using agent tools run into a fancier version of the same mess. Anthropic’s engineering write-up of a “think” tool is about agents that call other tools, not about your landlord. The useful bit is still the pause: “The “think” tool is for Claude, once it starts generating a response, to add a step to stop and think about whether it has all the information it needs to move forward.” You can steal the pause without the tool. After the first draft, ask: “Which of my facts did you use, and which did you invent?” If it invented the hardship story, you found the hole.
Do not paste passwords, account numbers, or a scan of your ID into a consumer chat to “give context.” The prompt can say “I already verified the account on the landlord portal” without the portal password.
Do I write a shorter prompt for a reasoning model?
Often yes. For a think-harder or o-series mode, give the goal, the done line, and the facts without prescribing every paragraph. Ordinary ChatGPT may still benefit from an outline. Choose the level of detail for the model you are using.
OpenAI’s prompting guide splits the job by model type: “Generally speaking, reasoning models will provide better results on tasks with only high-level guidance. This differs from GPT models, which benefit from very precise instructions.” On the ordinary ChatGPT or Gemini box, keep the steps. Those models still want the outline.
| What you are talking to | What to put in the prompt | What to leave out |
|---|---|---|
| Everyday ChatGPT / Gemini chat | Goal, audience, length, one example, facts | A ten-step inner monologue |
| A reasoning / “think longer” mode | Goal, done line, facts, stop rules | Micro-managing each paragraph |
| An agent that can use tools | When to look something up, when to ask you | “Search the web” on a task that is already in the paste |
Anthropic also reported a result from the airline domain of the customer-service benchmark τ-bench. With domain examples for a scratchpad tool in the prompt, the same Claude setup reached 0.570 pass^1 against a 0.370 baseline, a 54 percent relative gain; the model and tools stayed the same. A rent email needs no such tool, but a task-specific example can be more useful than a generic “reason step by step.”
| τ-bench airline, Anthropic report | pass^1 |
|---|---|
| Baseline prompt | 0.370 |
| Same model, prompt plus think-tool examples | 0.570 |
Both approaches need a stop rule: “If you do not know the late fee, ask. Do not guess a dollar amount.” A persona line such as “You are an expert property lawyer” does not prevent a fabricated statute.
This draft has no measured “best” word count for ChatGPT. Unverified. Count the jobs instead of the adjectives: ask for the body first, then the subject line if you need both. Two focused turns may work better than one stuffed prompt.
What else should I check if the prompt still fails?
Read the reply for invented facts before you argue with the tone. If the model added a sick pet, a dollar amount, or a law, your done line never banned fiction. Fix that, send once more, then stop rewriting the adjectives.
Should I start every prompt with “you are an expert”?
Only if you care about tone. OpenAI now treats a role line as a way to set voice and scope, not as a trick that makes the answer more correct. “You are a landlord-tenant lawyer in California” will make the draft sound legal. It will not check the Civil Code. For a two-sentence apology, skip the costume.
Is a longer prompt always better?
No. A 900-word instruction can produce a worse email than a 60-word one if it demands warmth, brevity, firmness, and a full timeline at once. Add the facts and one example instead. If the reply is still wrong, change one instruction rather than twelve.
Can I reuse the same prompt every week?
Yes, if the facts sit in a labeled block you swap. Keep “two sentences, Gmail, no legal advice” fixed. Replace the date and the amount. Google warns that mixed names for the same thing, like calling it a spreadsheet then a matrix, drag the chat down. Pick one word and stick to it.
What if I am not writing email?
The same skeleton holds for a meeting recap, a shopping list, or a product description. Job, audience, length, one sample, facts, and a ban on invented numbers. A listing that says “about 20 hours battery” when you never measured 20 hours is the landlord email again, just with a SKU.
