Lessons from writing with AI
Stop thinking of AI output as drafts to accept or reject. Instead, treat it as material to disassemble and recombine. The result: sharper writing
LET’S start with the uncomfortable admission: Writing – with all due deference to our editors – turned out to be something artificial intelligence (AI) could do.
It synthesised hundreds of academic papers. It built a computational model calibrated to Singapore’s labour market. It produced structurally sound, factually accurate and editorially lifeless prose.
Every draft hedged where conviction was warranted and treated all findings with equal weight. Competent mush. If you ship AI’s first draft, you’ve misunderstood the tool.
But that tool has changed faster than most people realise.
From prompting to partnership
When we started three months ago, we worked the way most people still describe AI collaboration: write a prompt, get a response, refine, repeat.
That model was obsolete within weeks – not because we got better at prompting, but because the tools shifted under us.
We began working with agentic AI – tools such as Claude Code that operate more like a research partner, with access to your full project, than a chatbot waiting for instructions.
Instead of asking AI to “summarise this paper”, we could point it at 30 papers, our working outline and a previous draft, and say: “The expertise section isn’t earning its place. Here’s why. Rework it using Brynjolfsson’s evidence, but keep the tone closer to Draft Three than Draft Five.”
It understood project context, remembered earlier decisions, and built on what existed rather than starting afresh.
Variations as building blocks
The most useful capability was not generating a good draft. It was generating variations – multiple versions of the same argument with different framings, structures and emphasis.
Our central argument started as the “expertise paradox” – AI destroys the training ground for the judgment it makes more valuable.
Six drafts in, something felt off. The framing was anxious, deficit-oriented. We asked AI for variations: one leading with what humans gain rather than what they lose, one built around partnership rather than paradox, one structured as opportunity rather than warning.
But here’s the important point: the insight that AI clarifies what was always most valuable about human work – that came from us. We knew what we believed.
The variations didn’t generate our argument; they helped us find the sharpest expression of it. We’d been writing about a problem. The better piece was about a possibility. The AI didn’t see that. We did.
We stopped thinking of AI output as drafts to accept or reject, and started treating it as material to disassemble and recombine. A paragraph from this version. A transition from that one. A better articulation of something we already knew, but had not yet said clearly.
The editorial was assembled, not written – and nothing cost us enough to defend irrationally.
We argued that AI is clarifying what is most valuable about human work. Co-authoring it with AI proved the point.
Five things we’d tell someone starting today
- Know what you think first. AI amplifies signal. If there’s no signal, it amplifies noise. When our argument was clear, AI built around it brilliantly. When we were vague, it produced competent mush. How: Before opening any AI tool, write down your core claim in one sentence – by hand if necessary. If you can’t, you’re not ready for AI. You’re ready for a walk.
- Use agentic tools, not chatbots. The difference between prompting a chatbot and working with an agentic AI, is the difference between texting a colleague and sitting next to one who can see your screen. Context changes everything. How: Tools such as Claude Code, Cursor and similar agentic environments let AI access your full project – files, drafts, data, previous versions. Point it at the project, not a blank prompt window.
- Generate variations, not drafts. Don’t ask for one good version. Ask for three with different emphases, then take the best elements from each. You’re assembling, not approving. How: Provide explicit instructions. “Give me three versions of this section – one leading with the data, one leading with the anecdote, and one leading with the policy implication. Keep each under 300 words. Then cut and recombine.”
- Set up AI as your critic. The generate-critique-revise loop was our most productive workflow. AI has no social incentive to be polite, which makes it a surprisingly effective editor. How: Give AI the working draft and a role. “You are a sceptical editor of The Business Times on a tight day. Where is this vague? Where does it lose the reader? Where does it assume too much?” Use the critique to guide revision – sometimes with AI, sometimes by hand.
- Throw most of it away. AI synthesised over a hundred papers, built models and generated frameworks. Much of it was good work. Most of it didn’t belong in the final piece. AI expands possibilities. Judgment narrows them to what matters. That narrowing is where the value lives. How: Use AI for compression directly: “Take this 3,000-word section and give me three versions at 500 words, each preserving a different core argument.” Pick the best, restore what was lost, compress again.
Taimur Baig is chief economist at DBS; Leslie Teo is senior director at AI Singapore and a member of the United Nations Independent International Scientific Panel on AI.
This is the companion piece to the commentary, “To boldly go: redefining jobs and expertise in the AI era”, which the writers co-authored with large language models.
TRENDING NOW
If Europe goes to war, the fallout will be felt in East Asia
China’s market is tougher than ever. Where can Singapore firms still find room to grow?
Vietnam’s race for growth tests pace of energy transition: experts
Indonesia stocks suffer worst plunge in nearly 30 years as MSCI proposes standalone status
