Tin Roofs of Surabaya
by Adit WijayaIn Surabaya in the late 1940s, a kretek factory clerk finds her dead brother’s name on a payroll list weeks after his burial.
When people say AI is disappointing, they often mean something narrower: they asked for a draft and got a pile of polished generalities. The wording looked competent, but it did not sound like their workplace, did not fit the situation, and did not save much time. They still had to rewrite it from scratch.
This is usually not because the tool is useless. It is because the request was too empty. Office writing depends on context. A client update email is not just “an email”. It has a purpose, a relationship, a level of detail, a risk level, and a next step. If you leave those out, the model fills the gaps with average office language. Average office language is what creates sludge.
That is why better prompting is less about clever phrasing and more about supplying the missing context. For most drafting tasks, six ingredients matter:
None of this is fashionable prompt engineering. It is ordinary briefing. If a colleague asked you for help writing a client message, you would not say, “Write an update email,” and walk away. You would explain what happened, who the client is, what tone is needed, what can be promised, and what should not be said yet. AI needs the same kind of briefing, often more explicitly.
This chapter focuses on usable first drafts. It does not change the rule from What Stays With You: you still keep judgment, accountability, and review. The goal is simpler. You want the first draft to be specific enough that editing is easier than starting over.
You do not need a rigid template for every request, but you do need to cover the basics. If a draft comes back weak, one of these pieces is usually missing.
Role helps narrow the style and purpose. “Act as a project coordinator writing to a client” is more useful than no role at all. It does not mean the model becomes accountable for the work. It simply gives the draft a lane.
Task should be concrete. “Draft a client update email” is better than “help with communication”. If you want options, say so. If you want a rewrite rather than fresh writing, say that instead.
Audience matters because office writing changes with the reader. A note to a teammate can be blunt and operational. A note to a client may need reassurance, careful wording, and a clear next step. If the audience includes senior people, legal risk, or strained relationships, that affects what AI should and should not draft.
Source material is where many prompts fail. If you ask the model to write from nothing, it will invent a plausible shape for the message. That shape may be wrong. Give it the facts you already have, even if they are rough. Later, in Give the Model Something to Work With, we will look at this in more detail. For now, the rule is simple: if the facts matter, provide them.
Constraints turn generic writing into work writing. These include length, tone, banned phrases, required points, things not yet approved, and what not to overstate. “Do not apologise for delay beyond what is stated” is a constraint. “Keep it under 150 words” is a constraint. “Do not mention the vendor issue” is a constraint.
Output format saves time at the end. If you want a subject line and a short email body, ask for that. If you want three versions with different tones, ask for that. If you want bullets you can paste into a slide, ask for bullets.
Put together, these pieces move the model from guessing to shaping. That is the difference between sludge and a draft you can actually use.
Take a common task. Priya works in account management. She needs to send a client update after a software rollout slipped by a week. The client already knows there were issues, but the relationship is still good. Priya wants to acknowledge the delay, explain what is happening now, and set expectations without sounding evasive or dramatic.
Her first prompt is the kind many people try:
Write a client update email about a delayed project.
The result is smooth, but empty:
Subject: Project Update
Dear Client,
I hope you are well. I wanted to provide you with an update on the project. We have experienced some delays due to unforeseen circumstances, but our team is working hard to address them. We remain committed to delivering a high-quality outcome and appreciate your patience and understanding during this time.
We will continue to keep you informed and will share further updates as they become available.
Best regards,
This is not wrong in a grammatical sense. It is wrong in an office sense. It says almost nothing. It hides behind stock phrases. It does not tell the client what changed, what is happening now, when they will hear next, or what Priya wants them to do. It could fit a hundred situations, which means it fits none of them well.
The problem is not that the model “is bad at email”. The prompt gave it no usable context. It had to invent a safe corporate shape, and safe corporate shape is usually the last thing you need.
Before improving the prompt, notice something practical: this is already a clue about your task. If the tool can only produce bland filler from your request, your request is under-specified. Do not keep asking for “better” in the abstract. Add information.
Priya tries again, this time adding audience and purpose:
Draft a short client update email to a customer about a one-week delay to a software rollout. The goal is to acknowledge the delay, explain that testing found issues, and reassure them that the team is working on it.
This is better, but still likely to produce generic reassurance. It may mention testing issues, but it still lacks enough detail to sound grounded. The model does not know what has happened, what has been approved for communication, or what tone suits the relationship.
So Priya adds source material and constraints:
Draft a client update email to our client, Northbridge, about a one-week delay to the payroll reporting rollout. Use a calm, direct tone. Keep it under 170 words. The client already knows we found issues in final testing last week. Facts to use: the issue affects report export formatting, no payroll data has been lost, the core system is working, we are retesting after a fix, and we now expect rollout next Thursday if testing completes as planned. Include a clear next update date of Tuesday. Do not over-apologise or make promises beyond next Thursday being the current target.
Now the model has material to work with. A likely draft is already more useful because it contains actual content and a boundary around what can be said. But Priya can still improve the odds by specifying format and what kind of wording to avoid:
Draft a client update email to Northbridge about a one-week delay to the payroll reporting rollout.
Audience: external client contact who is practical and prefers concise updates.
Purpose: acknowledge the delay, explain the current status, maintain confidence, and set the next communication point.
Facts to use and not go beyond:
- In final testing last week we found an issue affecting report export formatting.
- No payroll data has been lost.
- The core system is working as expected.
- A fix has been applied and the team is retesting now.
- If testing completes as planned, the new target rollout date is next Thursday, 14 March.
- We will send the next update on Tuesday, 12 March.
Constraints:
- Keep it between 120 and 160 words.
- Use plain business English, not inflated corporate language.
- Do not say “unforeseen circumstances”, “working hard”, or “thank you for your patience”.
- Do not sound defensive.
- Do not imply the date is guaranteed.
Output: provide a subject line and one email draft.
At this point the model is much more likely to produce something like this:
Subject: Update on payroll reporting rollout
Hi Sarah,
A quick update on the payroll reporting rollout. During final testing last week, we identified an issue affecting report export formatting. The core system is operating as expected, and no payroll data has been lost.
We have applied a fix and are retesting now. If testing completes as planned, our current target is to roll out on Thursday 14 March.
I will send you a further update on Tuesday 12 March, or sooner if anything changes before then.
Best,
Priya
This is not final just because it is better. Priya still needs to decide whether “A quick update” sounds right, whether to mention impact on users, and whether the date format matches the client’s norm. But now she is editing a specific draft instead of excavating one from sludge.
The step-by-step improvement matters because it shows what changed:
That is the pattern to copy.
Once you have a decent first draft, the next mistake is to throw it away and start over with another broad request. It is usually faster to steer the existing draft in small directions.
For example, Priya could follow up with:
These are useful because they preserve the source material and tighten one variable at a time. You are not asking the model to rediscover the task. You are directing edits.
This is also where your workplace voice starts to matter. Later, in Write in Your Workplace Voice, we will look at tone in more detail. For now, the practical point is that “professional” is too vague to be a good instruction. Better constraints are things like:
One more useful habit is to ask for options only when the choice is genuinely helpful. “Give me three versions” can be useful if you need to compare tone. It is wasteful if you already know what the message should roughly sound like. More output is not always better output.
And if you find yourself writing a very long prompt for a very simple task, pause. The point is not to perform complexity. The point is to supply the context the model cannot infer safely. A two-line prompt can be enough if the task is straightforward and the source material is clear. A longer prompt is only justified when the stakes, audience, or constraints make ambiguity expensive.
Pick one ordinary writing task you do often: a client update, a project recap, a status email, or a meeting follow-up. Do not choose a sensitive people issue or anything confidential. Write the first prompt the lazy way, then improve it by adding the six pieces from this chapter: role, task, audience, source material, constraints, and output format.
Compare the two drafts. Notice which extra detail actually changed the result and which did not. Keep the useful wording as a small prompt pattern you can reuse. The aim is not to become clever at prompting. It is to get to a usable first draft faster, with less rewriting and less guesswork.
Next, you will look at a related problem: not weak instructions, but weak input. Many bad drafts happen because the model is asked to invent what should have been supplied. The next chapter shows how to give it real material to transform.
A practical book for office workers who want AI to save time without creating more risk. It shows what to delegate, what to keep in human hands, and how to check drafts, summaries, analysis, and email before they leave your desk.
In Surabaya in the late 1940s, a kretek factory clerk finds her dead brother’s name on a payroll list weeks after his burial.
In the cliff town of San Jerónimo, Rosa Varela keeps the lighthouse lit with a trade everyone understands: a memory for a flame.
In a damp, overbuilt Newcastle a public service called Proxy lets people rent out their bodies for remote work, care visits and difficult conversations.
This book turns a steady paycheck into a system that covers bills, builds reserves, and leaves room for real life.
Sign in to leave a comment.