Writing With AI
Drafting, editing, and summarizing are three different jobs
① Using the same kind of prompt for drafting, editing, and summarizing is the shared cause of failure. The three need opposite amounts of latitude.
② Give drafting material, give editing a locked scope, give summarizing a discard rule. Those three sentences are the whole article.
③ Editing is the riskiest. AI is best at it — and while smoothing a sentence it deletes your hedges and conditions along with the roughness.
① Why "write something good" fails
If the seven prompting principles are the grammar that applies to every task, this article is that grammar applied to one domain: writing.
Start with an observation. When people ask AI to write, the request is usually one of three sentences.
- "Write me a blog post."
- "Clean this up."
- "Summarize this."
They look like completely different requests, but structurally they are identical. Each hands the model full authority over what "good" means here. So all three fail — and the catch is that they fail in different ways.
| Task | Failure signal | Cause | Fix |
|---|---|---|---|
| Draft | Nothing is wrong, but it's bland and says nothing | With no material, it fills in the training-data average | Supply material |
| Edit | Reads better, but your meaning shifted slightly | It fixes things you never asked it to touch | Lock the scope |
| Summary | Nothing is wrong, but the part that mattered is gone | The model decided what to drop | Set the discard rule |
That yields the key idea: latitude — how much discretion you hand the model. And the right amount runs in opposite directions across the three tasks.
Drafting wants high latitude, editing wants the minimum, and summarizing wants a middle setting where you specify only the direction. Yet most people use the same one-line prompt for all three. That is the entire problem.
② Drafting — supply material instead of asking it to invent
There is a structural reason AI drafts read as bland. An LLM picks the most plausible next token, and "most plausible" means "most average." Give it no material and it fills the blanks with the average treatment of that topic from its training data.
So "write about investing in US ETFs" returns something you feel you've read before. Not one sentence is wrong. There is also no reason the piece belongs to you.
| Prompt | |
|---|---|
| Weak | Write a blog post about investing in US ETFs, around 1,000 words. |
| Strong | Draft my blog post. Here is my material. · Reader: a professional in their 30s, three years of single-stock experience, new to ETFs · My claim: "it's diversified, so it's safe" is what makes people misunderstand ETFs · Facts I want used: ① index ETFs carry heavy top-holding concentration ② leveraged ETFs drift from the underlying index over long holds ③ expense-ratio gaps compound over time · My experience: I bought a semiconductor ETF thinking it was diversification, then watched it track single stocks almost exactly · Don't: recommend tickers, or close with "consult a professional" Use only this material to build the structure. Do not add facts I didn't give you. |
One caveat. The material itself has to come from source documents, not from the model. If you ask the AI to fill in facts like "top-holding concentration" or "long-hold drift in leveraged ETFs," you haven't supplied material — you've asked it to invent again. Facts of that kind belong in a source-based document like the ETF starter guide, checked there and carried over.
The difference isn't length — it's whether material exists. In the strong version the model isn't "writing"; it is arranging what you already have. That is the thing AI is genuinely good at.
The concrete way to raise latitude is to request more than one. "Write three openings from different angles — ① start from personal experience ② break a common misconception ③ lead with a single number." Pick one, or splice them.
But keep this to drafting. Multiple options in editing or summarizing only add decision load without raising quality, because those are tasks where exactly one answer should be right.
Finally, be precise about what a draft is for. An AI draft is not a finished piece; it's a way out of the blank page and a mirror for the structure of your own thinking. Reading it and noticing "this part is thin" is where its job ends. Publish it as-is and you start the problem covered in §6.
③ Editing — the task it does best and the one that's most dangerous
Honestly, editing is where AI is far and away strongest of the three. Typos, broken syntax, redundancy, sagging sentences, awkward particles — it catches in one pass what a person misses over several reads.
The problem is that it also fixes what you never asked about, and that overreach happens silently.
| What you wrote | What it "cleaned up" | What changed |
|---|---|---|
| Lower rates may favor growth stocks | Rate cuts favor growth stocks | Possibility became assertion |
| Some analyses raise oversupply concerns | Oversupply concerns are mounting | Attribution vanished; the subject blurred |
| On last quarter's numbers, it improved | Earnings improved | The time condition was deleted |
Look at what these share. The model did make each sentence more readable — shorter, better rhythm. And all three added confidence that wasn't in the original. Conditional clauses and hedges are exactly what makes a sentence sag, so they are the first casualties of "clean this up."
In investment writing that is not a small matter. The distance between "may favor" and "favors" isn't a question of style — it's a question of liability.
① Lock the scope — "Fix only spelling, grammar, broken syntax, and redundancy. Leave word choice, sentence structure, level of certainty, and argument as they are. Never delete conditionals or hedges ('may', 'some', 'as of')."
② Demand a change log — "List every change in a table: original / revised / reason."
The second matters most. Receive only a clean version and you have nothing to compare against, which makes review structurally impossible. In table form, overreach becomes visible.
One more rule: put numbers and proper nouns out of scope. If $20.3B becomes $23B the sentence still reads perfectly — and because it reads perfectly, nobody catches it. Why models are shaky with figures is covered in why AI gets numbers wrong. It's safer to keep a standing line in the prompt: "Leave all numbers, tickers, names, and dates exactly as written."
④ Summarizing — you decide what gets thrown away
"Summarize this" hides three delegations: what matters, who is reading, and how much to cut. Leave all three unset and the model applies its default — it treats whatever recurs and gets emphasized in the source as important. There is no reason that matches what's important to you.
Put precisely: summarizing is lossy compression. Losing information isn't the failure; it's the premise. The real failure is never deciding which direction to lose it in.
| Purpose | Must survive | Safe to drop |
|---|---|---|
| To decide (buy or not) | Conclusion, key evidence, counter-evidence, preconditions | Background, definitions, anecdotes |
| To share (explain to someone) | Context, why it matters, plain wording | Fine-grained figures, methodology |
| To file (find it later) | Numbers, dates, proper nouns, sources | Interpretation, commentary |
From one document, those three summaries are entirely different pieces of writing. Which is why a single line of purpose changes the output so much. For example: "Summarize this report for filing. Preserve every number, date, and company name; drop the analyst's interpretation and outlook and keep only the facts."
The editing problem shows up worse here. Because summarizing is by definition sentence-shortening, "under certain conditions X may happen" is structurally pushed toward "X happens." The caution in the original resurfaces as a prediction in the summary.
The only defense is an explicit instruction: "Keep every condition and assumption attached. Never convert something the source called a possibility into a fact."
For earnings materials or sell-side reports, this sequence holds up.
① Split facts from interpretation — "Table 1: only figures and dates stated in the source. Table 2: only the author's interpretation and outlook." Mixed together, the two become impossible to separate.
② Verify numbers against the source — never trust the figures in the summary itself. Knowing what to look for, as in the big-tech earnings guide, narrows where you have to check.
③ Chunk long documents — push in 100 pages at once and the middle goes hazy. The reason and the workaround are in how to make AI read your documents and tokens and the context window.
For the full arc from summary through verification, researching with AI lays it out in four stages.
⑤ The only way to make the voice yours
"Make it natural," "not too stiff," "friendly tone" — adjective instructions barely work. The reason is simple: "natural" means the training-data average to a model. You don't want the average; you want one specific voice, and an adjective cannot point at it.
Exactly one method works. Paste in things you wrote before.
"Below are excerpts from three pieces I wrote. Match the paragraph length, sentence rhythm, sentence endings, and how much jargon I use. Write new content, but keep my voice."
This is principle 4 (give examples) from the seven principles, applied to writing. Three or four paragraphs are enough, and more samples make it sharper.
Listing the habits you dislike works far better than describing the tone you want, because adjectives leave room for interpretation and a banned list doesn't.
For example: "Do not use — ① vague hedged endings like 'it seems that' ② subheadings that end in a question mark ③ 'in conclusion' in the final paragraph ④ 'not only A but also B' constructions ⑤ endings that state both sides and stop."
Build this list once and reuse it forever.
⑥ The fingerprint of AI prose — why not to ship it as-is
AI-written text has recurring tells. This isn't about detection tools. The practical problem is that readers find it boring.
- Paragraphs are unnaturally uniform in length — human writing mixes one-sentence paragraphs with long ones
- Every paragraph opens with its topic sentence — textbook-correct, rhythmically dead
- Too much parallel construction — "not only A but also B," "while simultaneously"
- Compulsive balance — every claim gets an equal-weight counterpoint until the piece says nothing
- The ending re-summarizes the body — making you read what you just read
The fix is mechanical. Delete one paragraph entirely, add a one-sentence paragraph, and cut the summarizing final paragraph. Those three moves alone change how the piece lands.
Ship AI sentences without review and any false statement inside them isn't the model's mistake; it's an error published under your name. Figures, years, and quotations the model slipped into a draft may well be hallucinations — and they pass review more easily precisely because the prose reads well.
One rule covers it: if a fact appears in the output that you didn't supply as material, verify it at the original source before publishing.
⑦ Prompts by situation
Applying the principles above to common situations gives the following. Rather than copying these verbatim, look at which element sits in which slot.
| Situation | Core instruction |
|---|---|
| Work email | Your relationship to the recipient, a one-line purpose, the action they should take, a length cap. "Background in two sentences max; requests as a list." |
| Report draft | Conclusion first. "Use only the data I gave you; mark missing items 'no data.' Do not fill gaps with estimates." |
| Editing your own text | Locked scope plus a change-log table. State that numbers and proper nouns are out of scope. |
| Long-document summary | Name the purpose (decide / share / file), require conditionals preserved, split fact from interpretation. |
| Social posts and short copy | Character cap, banned list, two examples. The shorter the piece, the more examples do the work. |
| Presentation script | Say it's meant to be spoken. "Should sound natural read aloud, sentences under 20 words, minimal connectives." |
Material for drafting, scope for editing, criteria for summarizing. Dropping the single habit of using one prompt for all three changes half the result on its own. Don't hand over the judgment — that's the whole point of this article.
Who holds the rights to what you produce this way — when an AI draft becomes your work — continues in who owns what AI writes.
※ Written as of August 2026, describing general principles. Behavior and default voice vary by AI service, and the sample prompts are meant to be adapted rather than copied verbatim. When handling documents that feed an investment decision, verify figures in the source rather than in a summary.
※ This guide is provided for general educational purposes and simplifies technical details for readability.
New guides, when they land
We publish AI literacy guides twice a week. Subscribe and the next one comes to you — free, unsubscribe anytime.