Chapter 03 · Section I · 14 min read
Writing and editing
The everyday text tasks the models handle well — drafting, summarising, rewriting, translating tone — and the practical habits that turn the model into a useful draft writer.
For most people in most jobs, the largest single use of generative AI is writing and editing text. Drafting emails. Summarising long documents. Rewriting paragraphs in a different tone. Generating reports from notes. These tasks are the quiet 80% of where the value of these tools is being captured.
This section is about doing them well. We focus on the practical habits that take the model from “a slightly impressive toy” to “a junior writer who never gets tired.”
Drafting
The most common use. You have a goal — an email, a memo, a one-page proposal, a project description — and you want a starting draft.
Best practice, in order:
- State the document type explicitly. “Draft a one-page project proposal.” Not just “write about this project.”
- Give it the structure you want. “Sections: Background (2 paragraphs), Approach (3 bullets), Timeline (table), Budget (table).”
- Paste in the raw material. Notes from a meeting, bullet points, an old version. The more context, the less the model invents.
- Specify the audience. “For a non-technical donor.” vs “For an engineering team.” changes everything.
- Ask for a first draft, not a final. Setting expectation that you will edit makes the model less precious with its words.
What you do not do: ask the model to invent the substantive content. The model doesn’t know what your project is about. Pasting your messy notes in and asking for a clean draft is the right move. Asking the model to imagine a project from scratch produces a generic blob that doesn’t sound like anything you would actually do.
Summarising
A nearly-magical capability, with one important caveat.
For factual summaries — meeting notes into bullet points, a 30-page report into a one-page executive summary, a long Nepali article into three paragraphs in English — modern models do this well, if the source material is in the prompt. Paste in the full text, ask for the summary, specify length and structure.
The caveat: when you ask the model to summarise something it has not been given directly — “summarise the latest news on the federal budget” — it is operating from memory of its training data, which may be out of date, partial, or invented. For factual summaries, always paste in the source.
A useful pattern for long documents:
Summarise this report in three sections: 1. The 3 most important findings (1 sentence each) 2. The recommendations (bulleted) 3. Open questions or things the report doesn’t address
Be specific. Quote numbers where they appear. Do not invent details.
Report: [paste text]
The “do not invent” line is not paranoia; it is calibration. It tells the model: I would rather a shorter, less polished summary than a beautifully fluent one that slips in invented details.
Rewriting and changing tone
A different kind of task — you have text, and you want a version of it. More formal. More casual. Shorter. Longer. In a different language.
This is one of the most reliable things modern models do. The output text is usually right, and easy to verify against the input.
Useful framings:
- “Rewrite this email to be one paragraph shorter without losing any key points.”
- “Translate this to formal Nepali suitable for a government letter.”
- “Make this paragraph sound less academic — like you’re explaining it to a friend.”
- “Convert this prose into three bullet points and a one-line summary.”
A pattern worth noting: paired rewrites are easier than open-ended ones. “Make this shorter” is vague; “rewrite this in 150 words instead of 400” is precise. Length, tone, audience, language — pick the dimension and be specific.
Translation, briefly
We will return to translation in section 3. But for text-only work, the basic capability is worth naming:
- Major languages (English ↔ Nepali, Hindi, common European languages). Near-professional quality from any frontier model. Better than Google Translate for nuanced texts; comparable or better for technical ones.
- Code-switched text (Nepali mixed with English, common in casual writing). Handled well by frontier models — better than dedicated translation tools, which sometimes choke on it.
- Low-resource Nepali languages (Maithili, Bhojpuri, Tharu). Inconsistent. Use with verification.
A useful trick: translate, then back-translate. Translate your English to Nepali, then ask the model to translate the Nepali back to English. If the back-translation matches your original meaning, the translation is solid. If it drifts, you’ve found the weak points.
Specific habits that help
Three small, high-leverage habits.
Read with the same care you’d give a human draft. Generative AI is good enough that bad drafts look correct at a glance. The mistakes are subtle — a small Nepali phrase that means something slightly different, a misremembered statistic, a missed nuance from your notes. Read with attention.
Keep your voice. If you use the model’s draft verbatim, your writing starts sounding like everyone else’s. The point is to edit toward your voice, not adopt the model’s default. Take what’s useful, change what isn’t yours.
Track what tasks it does well for you. Within a few weeks of using these tools, you’ll have a personal map: drafts of routine emails — great. Strategy documents — useful first draft, heavy editing. Anything that requires institutional knowledge — almost worthless. Knowing where the tool earns its keep is half the value.
A worked example: meeting notes → email
You have 200 words of messy notes from a meeting with a client. You need to send them a follow-up email.
The bad prompt:
Write a follow-up email about my meeting with Aastha Trading.
The model invents details.
The good prompt:
Draft a follow-up email in English from me to a client (Aastha Trading) summarising our meeting. Use the notes below as the source — do not add information that isn’t in the notes. Tone: warm but professional. Length: ~150 words. Structure: 1) brief thank-you, 2) summary of what we agreed, 3) next steps with dates.
Notes: [paste 200 words of raw notes]
Output: a draft you can send after 1–2 minutes of polish. The cost of the prompt was 30 seconds.
Check your understanding
Quick check
—You ask a model to summarise a 30-page report you haven't pasted in, but only described. The risk is:
Quick check
—The most common reason people complain that “ChatGPT writes generic stuff” is:
What comes next
We’ve covered drafting and editing. The next section is the other big text capability — analysis and extraction. Pulling structure out of messy text, classifying documents, turning notes into spreadsheets. These are the tasks where a model is often dramatically faster than a human.