LIVE: How Professional Writers Write with AI | Write-along

Every’s live write-along follows writer Katie Parrot and editor Jack through parts of their work on the same essay. Katie uses a rich set of folders, guides and review personas; Jack describes himself as an ultralight hiker who keeps his editing setup minimal. Their shared method is not a single perfect prompt. It is a sequence of decisions that each person still owns.

The session is also candid about its limits. A third participant cannot join. Jack edits a prepared draft rather than the exact introduction Katie writes live, and some ordinary checks are described or assumed to save time. The result is a demonstration of a process, not a complete article taken from idea to publication on screen.

That distinction makes the examples more useful. A viewer can study where the people intervene, which files or comments inform the model and what remains unresolved, without treating the live output as proof that all editorial work has already been done.

Katie keeps separate drafting folders for different columns. In style.md she defines the column, its purpose, its reader payoff and the kind of structure that tends to work. Her Working Overtime column connects first-person experiments with ways readers can understand their own work with AI.

One pattern is a friction-to-framework arc: begin with a personal difficulty, develop the example and arrive at an idea the reader can use. Katie says the pattern emerged when she asked a model to examine her existing essays. It describes her work, rather than prescribing what every essay must become.

A separate voice.md records sentence-level tendencies. Separating a column’s purpose from the way its sentences move gives the model different kinds of guidance. It also makes later problems easier to inspect than one undifferentiated instruction file.

To begin the essay, Katie asks the model to interview her one question at a time. She speaks about an incident in which her trusted writing context stopped helping: drafts became crowded and flat, and more feedback failed to recover the shape she wanted.

The interview does not simply proceed on autopilot. When the model asks about overdependence, Katie redirects it. She wants to explore what happens when context previously worth trusting becomes unreliable, rather than let the conversation become a different psychological essay.

She also asks the model to restart despite finding an existing outline, because this is a demonstration from first principles. Voice input is a personal preference, not a required ingredient. The product she uses, Monologue, is part of Every’s own product family, a relevant affiliation when interpreting the enthusiastic recommendation.

Katie traces the failure to a well-intentioned request to save everything: outlines, experiments, decisions and feedback. Optional templates gradually fused into conditions that every draft appeared to have to satisfy. A story, cultural connection, framework and joke all became competing obligations.

More instructions under deadline made the conflict worse. Her eventual response was to move old folders into a historical folder and rebuild the active examples, style guide and voice guide. The account is about separating useful working context from an archive, not destroying the historical record.

The practical distinction becomes explicit later in Jack’s edit: a record of an experiment should not silently become a rule for all future drafts. Katie now saves selected checkpoints deliberately. Her reported recovery is personal experience, not a controlled test proving that all large context collections harm writing.

Katie uses a 10% frame and a 30% outline, names inherited from her earlier content-marketing work. The first exposes the working thesis and story beats; the second develops the structure. The percentages name checkpoints rather than measure how much of the work is mathematically complete.

Her reasoning is that a weak premise becomes expensive to repair once a complete polished draft exists. She reads the thesis and considers the reader’s payoff before committing to the longer outline. If she already has an opening she likes, she supplies it herself; AI does not need to create every part.

The outline also reveals unresolved decisions. Should the reader see the consequences earlier? Is there too much material for the intended length? These are questions for the writer to settle while changes are still relatively cheap, not details the model should quietly decide on her behalf.

Katie’s reviewers examine specific qualities: a cold reader notices missing context, a suspense lens asks where tension dissipates, and other reviewers consider pace or objections. Some use names associated with admired storytellers, but Katie describes the results as AI reconstructions of principles, not the judgment of those actual writers.

The live example shows a useful disagreement. A reader report wants to reveal the damage earlier; a suspense-oriented report argues that showing the crash too soon would turn suspense into surprise. The model itself leaves the choice to Katie.

She treats the reports as prompts for reflection: do I agree, and is there something here I would otherwise miss? Running more personas does not remove that judgment. It can make competing editorial values visible, which is helpful only if someone decides which matters for this piece.

Katie resists the temptation to generate the whole essay at once. Even after approving an outline, she expects drafting to reveal discoveries that change the plan. She begins with the introduction, checking the hook, bridge, thesis and promise to the reader.

When the model produces an opening, she reacts at sentence level. A repetition it avoided may actually be useful; a line may sound natural but still need credit to the person who supplied the idea. The draft gives her something concrete to accept, reshape or reject.

She then decides which version deserves saving and moves useful text into Google Docs for manual work. The point is not that section-by-section generation guarantees quality. It preserves opportunities to notice a changed argument before a whole essay inherits the same mistake.

Jack says his normal first pass is to read the entire draft for big-picture questions. The livestream explicitly assumes that step for time, then demonstrates comments on the prepared version. He usually leaves the title and subtitle until he understands the piece deeply.

His comments often describe a problem without immediately solving it: what does “right” mean here, why would a new reader understand this phrase, and can this be said more directly? A personal detail may require Katie’s own experience rather than a model’s plausible replacement.

He also resists treating “sounds like AI” as a mechanical rule. A familiar AI-associated word can remain if it works in context. The editing question is what the sentence does for this reader, not whether a word list can prove who wrote it.

Jack uses a browser-based assistant to read comments and suggest revisions in chat. Selecting a passage helps scope the request to the text and its attached comment. When a suggestion improves one aspect but weakens another, he asks for more variations.

He reviews the options individually and decides what to suggest or apply. Some issues go back to Katie, particularly when the missing information is personal. The model is useful for exploring wording, but cannot supply her experience simply because a comment asks for specificity.

The familiar editing environment matters too. Jack values Google Docs’ undo behavior and comments, and does not want every activity moved into an orchestration interface. The tools shown are the team’s choices at the time of recording; the session is not a verification of today’s model availability or connector behavior.

Jack’s tightening skill aims for roughly a 10–15% reduction when the draft supports it, then stops before damaging meaning, voice, rhythm or necessary context. It first reads the whole draft and looks for structural bloat before shortening individual sentences.

82:00View full image

The examples favor concrete details, visible subjects and active verbs over abstract labels and throat-clearing transitions. A separate de-jargon skill translates technical language into plain language, tightens the result and then rereads the surrounding text for flow.

That last pass matters because a locally clearer sentence can repeat a nearby idea or disrupt a transition. Word-count reduction is therefore subordinate to the reading experience. A shorter version is not automatically the better edit.

After writer and editor revisions, a proofreader can catch production defects such as line breaks introduced while copying into a CMS. Jack also describes Katebench, an attempt to encode Kate’s recurring editorial preferences so obvious issues can be caught before her final pass.

Its suggestions enter the document as tracked changes. Jack accepts clear improvements, rejects clear mistakes and leaves uncertain cases for Kate. The goal is to free her attention for larger questions about standards and structure, not to declare that a simulation of her taste replaces her review.

The closing discussion calls this a test case, with a version of the article still to come. That is consistent with the process shown: AI helps expose options and perform scoped edits, while people retain the responsibility to verify context, supply experience and decide when the piece is ready.