How I AI: Farhad Manjoo's AI-Powered Writing Workflow for Research, Wordcraft, and Editing

In this episode of How I AI, former New York Times columnist Farhad Manjoo uses ChatGPT for web research, word choice, and structural feedback while keeping the reporting and final prose his own.

Farhad began experimenting when early models were still poor writers. As web search and language quality improved, ChatGPT moved from an occasional reference tool to a second window beside nearly every draft.

His workflow has three distinct jobs: map a research question, explore the nuance of a word or phrase, and critique the opening of a draft. He does not ask the model to write the article for him.

Map a research question

The AI research-assistant workflow mirrors Farhad's early-stage conversation: explore a possible claim, identify competing explanations, ask for sources, and decide whether the idea is strong enough to deserve reporting.

Farhad uses web-enabled ChatGPT to get oriented before he commits to an angle. The model accelerates discovery, but the linked sources remain the evidence.

Research a possible argument

The example begins with a broad question about commentary on tariffs.

Ask a conceptual question instead of a keyword query.

tell me about like, all the commentary on Trump's tariffs and, especially any that say the tariffs are good.

Read the synthesis as a map. In the demonstration, ChatGPT surfaced people, arguments, and links from outlets including Business Insider, the Detroit Free Press, and Reuters.

Narrow the inquiry with a follow-up question about a relevant industry or stakeholder.

is there anyone, in the automotive industry who has commented on the tariffs or,

Open the cited pages and confirm the author, date, context, and actual claim. A citation beside a generated sentence makes verification easier, but it does not guarantee that the synthesis is accurate or complete.

Farhad compared the experience to an always-available research assistant who can tolerate half-formed questions and rapid follow-ups. The advantage is speed and conversational exploration, not a substitute for reporting or source judgment.

Find a more precise word or phrase

The super-thesaurus workflow works through meaning and sentence context rather than returning a flat synonym list. Farhad can describe the shade of meaning he wants, reject suggestions, and test a candidate in the actual line.

Farhad uses ChatGPT as a conversational thesaurus. Instead of requesting a list in isolation, he supplies the sentence and explains the effect he wants.

Find an alternative idiom

One example started with "pay the piper," an idiom he felt was stale and not quite right for the sentence.

Give the model the full sentence.

For months, the mayor ignored public outrage over the polluted lake. Eventually, he realized he had to pay the piper.

Describe the metaphorical relationship, not just a synonym category.

I want to say this in a catchier way, but also with some kind of metaphor that describes like paying for something or, that like your previous actions coming home to roost or something.

Evaluate the suggestions. The model offered familiar options and the less expected "The devil came to collect," which Farhad considered usable.

Test a word in context

The same method works at the word level because the model can group alternatives by tone and then test one inside the sentence.

1. Ask for alternatives to the target word.
2. Compare categories such as dramatic, colloquial, formal, or ironic rather than treating every synonym as interchangeable.
3. Put a candidate back into the sentence and ask what meaning it changes. Farhad tested "grief" as a replacement for "outrage."

Does this work? For months the mayor ignored public grief over the polluted lake.

ChatGPT noted that grief signals sorrow more than anger and offered revisions for either intention. The writer still chooses the meaning and should reject language that is clichéd, unoriginal, or unsupported by the reporting.

Use AI as a first reader

The always-on first-reader workflow asks what the draft appears to argue, where attention drops, and which transitions or assumptions confuse a reader before the piece goes to an editor.

Farhad asks for feedback while a draft is still forming rather than waiting until the article is complete.

Get structural feedback during the draft

The purpose is diagnosis, not a wholesale rewrite. He asks about the speed of the opening, the placement of the argument, and unnecessary commentary.

Write the first six or seven paragraphs in your own document.

Paste the section into ChatGPT and ask narrow questions about structure, pacing, and clarity.

Does this get my point across quickly enough? Is there a way you can suggest to get to this argument much quicker? Am I doing too much unnecessary commentary here?

Decide which feedback improves the piece, revise in the document, and continue drafting. The model is one reader, not the editor of record.

Farhad does not rely on it to find deep logical inconsistencies. He finds it more useful for buried ledes, slow openings, unclear framing, and sections that drag. Claims, quotations, and source interpretation still require independent review.

The writer keeps the judgment

Farhad's method preserves a useful boundary: the model can widen the search, question a phrase, or react to a passage, while the writer owns the argument and every final sentence.

The most practical setup is simple: keep the draft and the AI conversation side by side, ask one specific question at a time, verify research against the linked source, and stop when the conversation begins to circle.

Farhad does not ask the model to write the column. He uses it at the moments when a writer is most likely to stall alone: before reporting, when the argument is still foggy; inside a sentence, when the available word is close but wrong; and during revision, when familiarity makes it difficult to experience the draft as a reader.

Research suggestions remain leads. Farhad follows the source, checks whether it supports the claim, and decides which evidence belongs in the piece. A confident summary without a verifiable source is not research, and a plausible counterargument is not automatically the most important one.

The thesaurus conversation is similarly iterative. He can explain that a word is too formal, too broad, too positive, or rhythmically wrong in the sentence. The model offers another set, but Farhad hears the line and makes the choice. This preserves the craft instead of averaging it into generic prose.

As a first reader, AI is available before another person should have to be. It can reflect the structure back, identify a missing bridge, or say what it thinks the thesis is. That reaction helps Farhad revise, but the editor and writer still own the standards, factual judgment, and final language.

Farhad keeps the questions narrow enough that he can evaluate the answer. He is not asking whether the entire draft is good. He asks what argument the reader sees, whether a transition follows, which word carries the intended shade of meaning, or what evidence would change the claim. A specific question produces feedback he can accept, reject, or investigate.

The conversation is also disposable. Once a source has been checked, a sentence has found its word, or a structural problem is understood, the model has done its job. Farhad does not need the assistant to preserve a permanent theory of his voice. He needs a patient interlocutor during the parts of writing that benefit from another mind in the room.

That makes AI useful without making it the audience. Farhad is still writing for people, working with an editor, and deciding what deserves their attention. The model helps him see the draft from another angle before that human relationship begins.